fix(deployment): join background deploy thread to stop cross-test state leak - #361

Summary

Fixes the intermittent CI failure

FAILED tests/unit/test_regressions.py::TestREG008NonInteractiveEnvDelete::test_undeploy_resource_force_remove_no_tty
  - _pickle.PicklingError: args[0] from __newobj__ args has the wrong class

which has been hitting a single Python version at random on unrelated PRs.

FAILED <nodeid> - <exc> means the test itself raised, not xdist failing to serialize a report — the victim test is innocent.

Root cause

1. deploy_all_background() spawns a daemon thread and returns nothing, so callers have no handle to join.
2. test_deploy_all_background patched get_or_deploy_resource, started the thread, and returned immediately. The patch unwound while the worker was still running, so the thread went on to hit the real deploy path with AsyncMock resources and cached them via _add_resource.
3. ResourceManager._resources is a class variable. The late write lands in whichever dict is current when it happens — one belonging to a later test.
4. The next test to call _save_resources() cloudpickles that leftover AsyncMock and dies.

The autouse reset_singletons fixture is already correct and function-scoped; it cannot help, because the rogue write happens after the reset. Which test pays depends on thread scheduling and xdist load distribution — hence the randomness.

Traced rather than inferred, via a plugin recording which test inserted each _resources entry:

### SAVE FAILED during test: tests/unit/test_regressions.py::TestREG008...
### CULPRIT key=<MagicMock name='mock.get_resource_key()'> type=<class 'unittest.mock.AsyncMock'>
    inserted_by=tests/unit/test_deployment.py::TestDeploymentOrchestrator::test_deploy_all_background

Changes

• deploy_all_background returns Optional[threading.Thread]. Backwards-compatible — it previously returned None implicitly, and no production caller reads the value.
• test_deploy_all_background joins inside the patch context, so the worker can no longer outlive the mock.
• New test_deploy_all_background_returns_joinable_thread pins the contract and asserts the deploy mock absorbed every resource.
• Empty-list path asserts None.
• Separate commit resyncs uv.lock's requires-python (<3.13 → <3.14) with pyproject.toml; CI already runs a 3.13 job. Resolution unchanged, no package versions move.

Test plan

Written test-first; the new test failed with AttributeError: 'NoneType' object has no attribute 'join' before the fix, and the pre-fix run visibly logged the leak (caching for cleanup ×3), now absent.

• The serial run reproduces the failure deterministically on main: 1 failed, 2629 passed → now 2618 passed, 0 failed
• make quality-check green (format, lint, 2618 parallel + 53 serial)
• mypy src/runpod_flash/core/deployment.py clean
• 3 consecutive -n auto runs, all 2618 passed