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.

Working directory path validation

Ensure that a repository's path (at initialization or open time) is
valid.  On Windows systems, this means that the longest known path
beneath the repository will fit within MAX_PATH: this is a lock file for
a loose object within the repository itself.

Other paths, like a very long loose reference, may fail to be opened
after the repository is opened.  These variable length paths will be
checked when they are accessed themselves.  This new functionality is
done at open to prevent needlessly checking every file in the gitdir
(eg, `MERGE_HEAD`) for its length when we could instead check once at
repository open time.

Add a simple accessor for workdir paths to get an absolute on-disk path
given a repository and a relative path within it.  This is useful to
avoid copy-pasta `git_buf_joinpath` and to ensure that we validate
working directory paths while honoring `core.longpaths` settings.

repo: remove an inappropriate use of PASSTHROUGH

We don't want the default branch's refname here but its name as branch.
Returning an error saying it's not a valid reference here suggests we want the
value of `init.defaultbranch` to be something like `refs/heads/default` which is
not the case.

This error code is for callbacks where we should act as though the callback was
not set. This is not something that makes sense for a `_foreach` and checking
for that error message to bubble up mostly is happenstance precisely because
this is not an error code we expect in the callback.

As part of removing this, let's also remove a use of foreach as we can open-code
this check.

We already do this in repo_init_head