test: stop a background writer failing teardown's rm

[ttncode]

What this changes

tests/compose.bats has been failing on CI in teardown, not in any
assertion:

not ok 19 install.sh generates a password for every service that has one
# (from function `teardown' in test file tests/compose.bats, line 9)
#   `rm -rf "$WORKDIR"' failed
# rm: cannot remove '/tmp/tmp.K0zZ6SlOXR/demo': Directory not empty

rm -rf reports "Directory not empty" when an entry appears after it has
already walked the directory — something is still writing into the generated
project while the test that generated it is being torn down. compose.bats
generates a project per test and the lane runs --jobs $(nproc), so on a
four-core runner several laravel generations overlap.

It is a race, not a defect in any test: a different test loses it each run.
Across five runs of the same job it took down {12, 19, 26}, then {19}, then
{20}.

It is also not new, and not this branch's. Measured by re-running the same job
on the same commits: main fails one run in three, and every not ok is this
same teardown.

Two changes:

• git config --global gc.auto 0 in the suite's own git configuration.
  scaffold new commits what it creates, and a commit can hand the repository
  to a detached git gc that outlives the command and keeps writing into
  .git. That is the likeliest writer this races. It was never caught in the
  act, so this is the suspect removed rather than the cause proven — a
  throwaway repository has nothing worth maintaining either way.
• remove_workdir replaces the bare rm -rf "$WORKDIR" in all thirteen
  teardowns. It retries for five seconds, and if the tree is genuinely stuck it
  still fails the test and prints what is left in it — so the next occurrence
  arrives as a diagnosis instead of another mystery.

No test's assertions change.

How it was verified

• bats tests/compose.bats — 17/17, the suite the failure lands in
• mise run test-unit — no failures
• mise run test-runner — both lanes
• shellcheck on tests/helpers/setup.bash — clean for the added function
• This PR's own CI

The retry cannot be proven to fire from here: the race has never reproduced
locally, across a full test-runner, three repeats of the individual test that
failed, and a probe that looked for processes holding the generated tree the
instant scaffold new returned. That is why the failure path reports rather
than just giving up.

Checklist

• mise run lint passes
• mise run test-runner passes
• New behaviour has a test that fails without the change — n/a, and left
  unticked rather than claimed: this changes test infrastructure, and the
  race it removes has never reproduced outside CI, so there is nothing to
  write a failing test against
• Docs that describe changed behaviour were updated in the same commit
• No unrelated changes