CDRIVER-4220 simplify and test obsolete bson_sync_synchronize API

[ghost]

The bson_sync_synchronize() API, a basic platform-specific wrapper around an unspecific memory fence, was added back in 2013 and used briefly for mongoc-counters. That use has since been replaced by bson_atomic_thread_fence().

We don't know of any other users of this API but it was technically exposed as public, so third party code may still rely on it appearing in bson-compat.h.

We plan to remove atomics from the API in 2.0, see:
#1762

To align bson_sync_synchronize() with our plans, this patch replaces the original macro definition with an inline function that's equivalent to bson_atomic_thread_fence(), and this function is marked as deprecated. We will remove the function entirely for 2.0, along with bson-atomic.h.

Manually tested that the expected deprecation appears when using bson_sync_synchronize() from external code. Included a trivial automated test which can't verify the deprecation warning or the memory fence functionality, but it will catch compile failures or crashes.

[eramongodb]

CC @vector-of-bool, requesting extra eyes to confirm this behavior is (sufficiently) equivalent to the former definition of bson_sync_synchronize given your prior work on bson-atomic.h. I think this should be acceptable (including platform support/compatibility) + I am in favor of avoiding/removing inline ASM where we can. Related GCC 4.8.4 documentation:

__sync_synchronize (...): This built-in function issues a full memory barrier.

[ghost]

The asm shouldn't be applicable on any supported platforms. Both the before and after implementations use __sync_synchronize() on gcc. I think in practice the biggest difference might be that unoptimized builds may see a function call now whereas previously that was impossible.

[vector-of-bool]

I think @mdbmes is correct, and the inline asm statements are no longer expanding on any supported platforms. I believe this change is good.

[vector-of-bool]

LGTM