fix: disable git auto-gc during dependency temp-dir clone+checkout

[mchen-sentry]

We're seeing intermittent acceptance failures on getsentry/sentry where devservices up blows up with shutil.Error from _checkout_dependency. Attempt 1 of run 25064611924 / job 73428191911:

File ".../devservices/utils/dependencies.py", line 662, in _checkout_dependency
    shutil.copytree(temp_dir, dst=dependency_repo_dir)
shutil.Error: [('/tmp/tmpoydmzhs4/.git/objects/info/commit-graphs/tmp_graph_8REK9p',
                '.../chartcuterie/.git/objects/info/commit-graphs/tmp_graph_8REK9p',
                "[Errno 2] No such file or directory: '/tmp/tmpoydmzhs4/.git/objects/info/commit-graphs/tmp_graph_8REK9p'")]

The same job logs git version 2.54.0, and 2.54.0's release notes call out that git maintenance now defaults to the geometric strategy. The new default writes more tmp_graph_* and commit-graph-chain.lock fragments via the create-then-rename pattern git uses for atomic commit-graph chain updates. After our git checkout returns, gc.auto fires in the background and starts writing those fragments while copytree is still walking the temp tree: os.scandir enumerates a tmp file, then shutil.copy2 hits ENOENT because git has already renamed it.

Fundamentally none of this background bookkeeping is useful inside a temp dir we're about to bulk-copy: any work either gets carried over verbatim during copytree or redone later when the cache is actually used. #312 patched a related race in TemporaryDirectory's implicit cleanup (the rmtree when the with block exits), but copytree doesn't go through that cleanup machinery so ignore_cleanup_errors=True doesn't apply here. We disable auto-gc on the two git invocations inside the temp dir (the partial clone and the checkout) by passing -c gc.auto=0 -c maintenance.auto=false -c fetch.writeCommitGraph=false. That stops the background gc from ever running against the temp dir, killing both this copytree race and the rmtree race #312 was patching around.

(Considered also adding ignore=shutil.ignore_patterns('tmp_*', '*.lock') to the copytree as belt-and-suspenders, but disabling gc at the source is the cleaner fix; the patterns approach would still have a residual race for any non-tmp file that maintenance might touch.)