FunctionalTests: make shared control-repo cache setup resilient

[tyrielv]

Problem

The GitCommands functional tests compare a GVFS repo against a plain "control" git repo. Every control repo fetches from one machine-global bare cache. That cache is set up in ControlGitRepo's static constructor, which checked Directory.Exists and then either cloned or fetched — and discarded every git exit code.

That setup is neither atomic nor verified. Fixtures run in parallel ([assembly: Parallelizable(ParallelScope.Fixtures)]) and the cache path is shared by concurrent test processes on the same machine, so a process can observe a half-built clone directory (Directory.Exists is true before the clone finishes) and fetch from an incomplete repo. A transient clone or fetch failure has the same effect. The cache is then left missing branches, the failure is swallowed, and every GitCommands test fails its setup checkout with:

error: pathspec 'FunctionalTests/20201014' did not match any file(s) known to git

This shows up as a large batch of GitCommands failures (e.g. 21 in one run), all originating in [SetUp] / CreateEnlistment, even though the product is fine.

Root cause, reproduced

A local harness that mirrors ControlGitRepo.Initialize():

• healthy cache → 0/50 control-repo builds fail
• cache missing the branch → 50/50 fail, with that exact error

So the failure is fully explained by an incomplete shared cache plus swallowed git errors.

Fix

GVFS.FunctionalTests/Tools/ControlGitRepo.cs:

• Serialize cache creation and refresh across processes with a system-wide mutex.
• Build the cache atomically: clone into a temp directory, verify the base branch is present, then move it into place, so no other process ever observes a partial cache.
• Retry transient clone and fetch failures, and rebuild the cache when verification fails.
• Fail loudly with a clear message when a control repo cannot fetch or check out its branch, and retry the whole control-repo build, instead of silently producing a broken repo that fails 20+ tests with a confusing cascade.

With the fix, a control repo that starts from a broken cache self-heals and 0/50 builds fail.

Scope

Test-infrastructure only; no shipped runtime behavior changes.

Verification

• Managed + native build clean (0 errors, 0 warnings).
• Fix harness starting from a broken cache self-heals → 0/50 fail.
• Full functional suite not run in this environment (needs the internal test repo, admin, and a ProjFS mount).