Memory optimization

Use CrtAllocator for stack.
ShrinkToFit stack after parsing.

Lazy allocation. (so Document not for parsing don't need to allocate
stack).
Remove redundant stackCapacity member variable.

.travis.yml: enable 32-bit builds

Use a tighter bound.

As Ubuntu supports running 32-bit executables based on its multiarch
setup, we can enable the tests on 32-bit configurations as well.

Currently, valgrind fails for 32-bit executables on the 64-bit host.
This requires some more investigation.

Move to RFC7159 (closes #90)

Fix sign-comparison warning/error on 32-bit platforms

By adding an optional CRTP template parameter, the BaseReaderHandler
can call the "overridden" `Default()` function from the `Derived`
class.

See https://en.wikipedia.org/wiki/Curiously_recurring_template_pattern

With the allowance of arbitrary root value types, the individual tests
can use the top-level Parse functions, instead of ParseFoo() variants.

Secondly, some unneeded array wrappers have been dropped and non-singular
tests starting with other values than objects or arrays have been added.

Fix #72

travis-doxygen.sh: only overwrite .git-credentials, iff not exists

It is simple and pure C++. And it is found in performance test that it
is even faster than the original version, due to distribution of n. But
the performance gain is not obvious in RapidJSON.

typo