bson_data() is rather slow when used in a hot path. Just call it once
and increment our buffer pointer manually. This is safe because we have
already done our buffer growth outside this hot path.

Let the compiler setup the constant mask directly.

This is actually a pretty healthy performance boost. We avoid
traversing the hierarchy which means we also don't pollute the cache
lines so much.

If we just initialize everything to the default, we can just use malloc
instead of calloc for a bit of a speed improvement.

The default pointer alignment is fine since we only care about cache
alignment for stack allocations using inline structures.

Various build system improvements.

 * Try to use -Bsymbolic to avoid PLT jumps.
 * Detect Windows (define OS_LINUX OS_WINDOWS)
 * Verify basic ABI compliance with abicheck.sh and libbson.symbols.
 * Cleanup bson/Makefile.include.

Libbson is pretty aggressive about warning the consumer of invalid API
use. However, that means that we short circuit some functions. This is
valid since consumers shipping invalid API use is nothing we can protect
against, but this will at least quiet some static analyzers.