There were some subtle semantic differences between the various
implementations of atomic functions. Now they behave the same, have
tests and are better documented to avoid this from happening again in
the future.

Of note:

* The semantics chosen for `git_atomic_compare_and_swap` match
  `InterlockedCompareExchangePointer`/`__sync_cal_compare_and_swap` now.
* The semantics chosen for `git_atomic_add` match
  `InterlockedAdd`/`__atomic_add_fetch`.
* `git_atomic_swap` and `git_atomic_load` still have a bit of semantic
  difference with the gcc builtins / msvc interlocked operations, since
  they require an l-value (not a pointer). If desired, this can be
  homogenized.