LIBZMQ-497 there could be unsent bytes in encoder

When we send a large message, the message can be splitted into two chunks.
One is in the encoder buffer and the other is the zero-copy pointer.
The session could get the term before the last chunk is sent.

Added IPv6 option to context

IPv6 related changes and cleanups to test cases

* Removed or truncated sleeps so the tests run faster
* Removed dependencies on zmq_utils
* Rewrote a few tests that were confusing
* Minor code cleanups

 LIBZMQ-498 - Remove heap allocations in zmq_poll for small poll item sets

Until now, zmq_poll always allocates the poll items on the heap.
Now, small item sets, up to ZMQ_POLLITEMS_DFLT, are stack allocated
and only larger sets are allocated on the heap.

Fix using wrong name for windows.h check with cmake

define a dummy EPROTO for platforms that don't use it

Fix perf-tools build on Linux with cmake

Adds tests/test_disconnect_inproc to .gitignore

Add bounds check on upstream XSUB messages

This accomplishes the same thing as 2e2ef7fd in a potentially cleaner way.

Ensures 0 length messages aren't being checked for subscribtion status on whatever data happens to be following the data pointer.

do not use EPROTO on platforms that do not have it

Shortening the =-chars to make asciidoc work

Minor autoconf tweaks

deprecate zmq_ctx_destroy in favor of zmq_ctx_term

AM_CONFIG_HEADER raises an 'obsolete error' with automake 1.13.

autotools warns that configure.in
should actually be named configure.ac

in order to avoid logical collisions with pre-existing notions of context destruction in bindings (czmq, pyzmq).

Clarified how icp:// works with zmq_bind

Fixed el5/el6 deps in RPM spec file / newline fix

Whitespace and comment cleanups