Our virtual commit must be the last argument to merge-base: since our
algorithm pushes _both_ parents of the virtual commit, it needs to be
the last argument, since merge-base:

> Given three commits A, B and C, git merge-base A B C will compute the
> merge base between A and a hypothetical commit M

We want to calculate the merge base between the actual commit ("two")
and the virtual commit ("one") - since one actually pushes its parents
to the merge-base calculation, we need to calculate the merge base of
"two" and the parents of one.

Virtual base building: ensure that the virtual base is created and
revwalked in the same way as git.

When the commits being merged have multiple merge bases, reverse the
order when creating the virtual merge base.  This is for compatibility
with git's merge-recursive algorithm, and ensures that we build
identical trees.

Git does this to try to use older merge bases first.  Per 8918b0c:

> It seems to be the only sane way to do it: when a two-head merge is
> done, and the merge-base and one of the two branches agree, the
> merge assumes that the other branch has something new.
>
> If we start creating virtual commits from newer merge-bases, and go
> back to older merge-bases, and then merge with newer commits again,
> chances are that a patch is lost, _because_ the merge-base and the
> head agree on it. Unlikely, yes, but it happened to me.

Test that shows that submodules are incorrectly considered in renames,
and `git_merge_trees` will fail to lookup the submodule as a blob.

When `GIT_MERGE_FIND_RENAMES` is set, provide a default for
`rename_threshold` when it is unset.

Since the `apply` callback can defer, the `check` callback is not
necessary.  Removing the `check` callback further makes the `payload`
unnecessary along with the `cleanup` callback.

Ensure that setting the merge attribute forces the built-in default
`text` driver and does *not* honor the `merge.default` configuration
option.  Further ensure that unsetting the merge attribute forces
a conflict (the `binary` driver).

When a `check` or `apply` callback function returns `GIT_EMERGECONFLICT`
stop and product a conflict.

When a `check` or `apply` callback function returns `GIT_PASSTHROUGH`,
move on to the default merge driver.

Consumers can now register custom merged drivers with
`git_merge_driver_register`.  This allows consumers to support the
merge drivers, as configured in `.gitattributes`.  Consumers will be
asked to perform the file-level merge when a custom driver is
configured.

Update unit test to use newfangled `st_ctime_nsec`, which provides
indirection to the platform-correct name.

Windows defines `timeval` with `long`, which we cannot
sanely cope with.  Instead, use a custom timeval struct.

When building a recursive merge base, allow conflicts to occur.
Use the file (with conflict markers) as the common ancestor.

The user has already seen and dealt with this conflict by virtue
of having a criss-cross merge.  If they resolved this conflict
identically in both branches, then there will be no conflict in the
result.  This is the best case scenario.

If they did not resolve the conflict identically in the two branches,
then we will generate a new conflict.  If the user is simply using
standard conflict output then the results will be fairly sensible.
But if the user is using a mergetool or using diff3 output, then the
common ancestor will be a conflict file (itself with diff3 output,
haha!).  This is quite terrible, but it matches git's behavior.

Add a simple recursive test - where multiple ancestors exist and
creating a virtual merge base from them would prevent a conflict.

Provide a new merge option, GIT_MERGE_TREE_FAIL_ON_CONFLICT, which
will stop on the first conflict and fail the merge operation with
GIT_EMERGECONFLICT.

xdiff craps the bed on large files.  Treat very large files as binary,
so that it doesn't even have to try.

Refactor our merge binary handling to better match git.git, which
looks for a NUL in the first 8000 bytes.

This allows us to remove OS checks from source code, instead relying
on CMake to detect whether or not `struct stat` has the nanoseconds
members we rely on.

The header src/cc-compat.h defines portable format specifiers PRIuZ, PRIdZ, and PRIxZ. The original report highlighted the need to use these specifiers in examples/network/fetch.c.  For this commit, I checked all C source and header files not in deps/ and transitioned to the appropriate format specifier where appropriate.

When invoked with three files that each lack a trailing newline,
the merge result should also lack a trailing newline.

As we attempt to replicate a situation in which an older checkout has
put a file on disk with different filtering settings from us, set the
timestamp on the entry and file to a second before we're performing the
operation so the entry in the index counts as old.

This way we can test that we're not looking at the on-disk file when the
index has the entry and we detect it as clean.