PackfileMaintenanceStepTests: fix flaky tests

[derrickstolee]

The PackfileMaintenanceStepTests have been failing periodically, and
mostly with failures in RepackAllToOnePack() leading to a failure in
ExpireAllButOneAndKeep().

The way we "force" Git to repack all is to provide a custom batch size.
This batch size used to mean "repack all pack-files with size at most
X until reaching a size of at least X". By specifying "total size - 1",
this required repacking all objects.

HOWEVER, the new logic actually computes an "expected" size of the
objects in a pack that are referenced by the multi-pack-index. If we
have duplicate objects, then those objects can cause two types of error:

1. The duplicates reduce the total expected size below the batch size.

2. The computation of "total expected size" causes round-off error.

Both types of error under-count the number of bytes in the final
pack-file.

By changing the batch size to "total size - min size + 1" we ensure the
following:

I. The batch size is larger than the maximum pack size.

II. Every pack-file is required to reach the batch size.

III. We have minimized the chance that duplicate objects will drop
below this threshold.

With the use of gvfs-helper, we have less control over the pack-files
that are downloaded as part of the clone operation. At minimum, with
this change we should be in a better state than before.

Signed-off-by: Derrick Stolee dstolee@microsoft.com

[derrickstolee]

See also microsoft/VFSForGit#1562.