Significantly reworked the monitoring infrastructure with a more granular per socket API and to play well with monitoring endpoints in application threads

Replaced device concept with proxy concept

* zmq_device is now a wrapper that calls zmq_proxy
* zmq_proxy adds capture socket

Add "#undef NDEBUG" to all tests.

This change makes sure that even if the tests are built in a
"release" configuration (with optimizations and NDEBUG turned on),
the assertions won't get compiled out of the tests themselves.

The C standard guarantees that the most recent inclusion of
<assert.h> is the one that counts, so it's important that the
"#undef NDEBUG/#include <assert.h>" come as the last thing in
the block of header files.

"testutil.hpp" includes <assert.h>, so I've left <assert.h> out
of any test that #includes "testutil.hpp", just for the sake of
brevity.

New message encoder/decoder + code simplification

This is supposed to become part of the ZMTP/1.1.

The main differences from the ZMTP/1.0 framing protocol are:
- flags field comes first, followed by the length field
- long messages are signaled using a flag rather then 0xff escape
- length field does not include the flags field, 0 is a valid value

Fix some PGM refactor breakages

Hopefully fixed LIBZMQ-427 - there was a slight typo in the init_address
refactor. The encoder refactoring had also broken pgm_sender and
receiver, but just required updating to use the new functions.

Code cleanup

Introduce versioning into ZTP

Since ZMQ 2.x does not support subscription forwarding, it's not
possible to use ZMQ 2.x SUB socket to receive messages from a PUB
socket.

This patch adds some compatibility layer so that ZMQ 2.x SUB socket
receives messages from PUB socket.

This patch introduces i_msg_sink and i_msg_source interfaces. This
allows us to make message encoder and decoder more general.

The new protocol adds support for protocol version and  exchanges the
socket type, so that the library can reject a connection when the
sockets do not match.

The protocol was designed so that it's possible to detect and fully
support ZTP/1.0 peers.

When a new connection is set up, peers exchange greeting messages. The
greeting message encodes both the protocol verion and the socket type.
The format of the greeting message is as follows:

    greeting    = tag1, adaptation, tag2, version, length, socket_type
    tag1        = BYTE / 0xff
    adaptation  = 8 BYTES
    tag2        = BYTE / 0x7f
    version     = BYTE / 1
    length      = BYTE / 1
    socket_type = BYTE

The protocol does not define the value of adaptation field.

When interoperability with ZTP/1.0 peers is required, the adaptaion
encodes, in network byte order, the length of identity message increased
by 1. When adaptaion consists of eight zeros, the current
implementatatio of 0MQ 2.x closes the connection.

This patch supports both ZTP/1.0 and new protocol.

Fixed COPYING.LESSER

Silence all "unused parameter" warnings from Clang.

This formerly unused parameter actually represents the socket
on which the event was received. As such, we should check that
its value makes sense: it must be either "rep" or "req", and in
the case of some kinds of events, it must be specifically one
or the other.

After this change, "s" is no longer unused.

Compiling without warnings is a good goal, because it makes
new warnings (which probably indicate bugs) stand out rather
than getting lost in the spam.

My fixes fall into two categories:

    - Adding (void) casts of unused parameters, where their
      unusedness seems like a TODO (or in some cases a bug?).

    - Removing parameter names altogether, where the function
      is clearly a stub that will never use its parameters.

Should be no change in behavior.

char_traits<unsigned char>::to_char_type(x) used to return 0 no matter
what x was, and likewise to_int_type(x) used to return 0 no matter what.
(0 is what you get when you default-construct an integral type, which
is what the old code was doing.) This seemed like buggy behavior to me,
so I've changed it.

Fix build breaking typos in monitor VA code

Include a note about potential interactions with reactors, and update
reference to old FAIL_UNROUTABLE name.

The endpoint was excluding the tcp://, causing tests to fail

Code as stands breaks shutdown process. It was a bugfix by Arthur to a
bad line which was testing for an impossible state - but afaics we do
actually want to flush in those states. It is possible I am wrong on
that though - if there are any shutdown issues introduced aroudn this
commit I would suggest further investigation around this flushing
behavior.

VA specific monitor event in socket base was named monitor_event rather
than va_monitor_event

The call to the va version of the function was using 'event' instead of
'event_'

Compile was failing on the monitor_event function function due to the
va_start params being switched.

Various bugfixes found by compiling with extra warnings

There are three versions of monitor_event(), all taking
variadic arguments. The original code just has the first one
creating a va_list and passing that va_list variadically to
the second one... which creates a new va_list and passes it
variadically to the third one... and of course everything
blows up when we try to pull a non-va_list argument off the
stack.

The correct approach matches the C standard library's use
of printf/vprintf, scanf/vscanf, and so on. Once you make
a va_list, you must pass it only to functions which expect
a va_list parameter.

Static analysis says:
src\tcp_address.cpp(297): error V595: The 'res' pointer was utilized before it was verified against nullptr. Check lines: 297, 301.
src\tcp_address.cpp(603): error V106: Implicit type conversion third argument 'full_bytes' of function 'memcmp' to memsize type.
src\tcp_address.cpp(603): error V526: The 'memcmp' function returns 0 if corresponding buffers are equal. Consider examining the condition for mistakes.

In fact the use of "memcmp" is correct, but the enclosing "if" isn't
necessary, and the compiler is happier if "full_bytes" is a size_t.

Static analysis says:
src\pipe.cpp(193): error V547: Expression is always false. Probably the '||' operator should be used here.

If flush() is called on a pipe whose state was
"terminated" or "double_terminated", the programmer's
intent was to return immediately. But in fact the
two conditions can never be true simultaneously, so
the early return never happens, and we may try to flush
a terminated pipe anyway.

Static analysis says:
src\fd.hpp(38): error V103: Implicit type conversion from memsize to 32-bit type.

Adding the explicit cast should shut it up.

Static analysis says:
src\zmq.cpp(489): error V220: Suspicious sequence of types castings: memsize -> 32-bit integer -> memsize. The value being casted: '* count_'.
src\zmq.cpp(510): error V127: An overflow of the 32-bit 'nread' variable is possible inside a long cycle which utilizes a memsize-type loop counter.

I've silenced the warning on line 489 and ignored the other.
But also, it looks to me like there's a serious bug here: The
out-parameter "count_" is never set to zero before we start
incrementing it. So its final value will always be between
1 and 2 times its initial value. The fix seems obvious.

[LIBZMQ-414] Fix an error in the inline assembly for Thumb2.