fix(testing): run prefect_test_harness teardown on exception-path exits

[zzstoatzz]

Scope

This PR does NOT fix the test_prefect_test_harness py3.14 postgres flake (#21405, #21414, #21460). It addresses a related-but-separate gap in prefect_test_harness exception handling that I noticed while investigating that flake.

The gap

prefect_test_harness is a @contextmanager that places test_server.stop() and run_coro_as_sync(drain_workers()) after the yield with no try/finally:

```python
yield
run_coro_as_sync(drain_workers())
test_server.stop()
```

If any exception propagates out of the harness body, the lines after `yield` are never executed. `test_server.stop()` is skipped, which means the `SubprocessASGIServer` singleton cleanup path (that #21405 carefully set up via `_instance_key`) is never reached — `_instances[None]` is left holding a stale entry with `running=True`, and the next `prefect_test_harness()` in the same process silently reuses it.

This is narrower than what the py3.14 flake exercises. The failing test (`test_prefect_test_harness`) runs clean — no exception — so this gap is invisible to that test. But any other test that raises from inside the harness would leave the process in a polluted state for subsequent tests, and that pollution is hard to attribute back to the source. I hit this gap by accident while writing a test that raised inside the harness and saw the next harness start() become a no-op.

Fix

Register `test_server.stop()` and the log/event drain on the harness's existing `ExitStack` so both run regardless of how the harness body exits. LIFO ordering means the drain runs while the server is still alive (before `stop()`), matching the prior ordering.

```python
test_server = SubprocessASGIServer()
stack.callback(test_server.stop) # runs on unwind, LIFO
test_server.start(...)
stack.enter_context(temporary_settings({PREFECT_API_URL: test_server.api_url}))
...
stack.callback(lambda: run_coro_as_sync(drain_workers())) # runs first (LIFO)
yield
```

What this deliberately does NOT include

• The `SubprocessASGIServer.start()` "dead subprocess" health check that was in the earlier draft of this PR. It was speculative defense-in-depth, not tied to any observed failure. The py3.14 flake's stale subprocess is alive, just pointed at the wrong database, so the health check doesn't help it. Dropped to keep this PR focused.
• Any attempt to fix the py3.14 postgres flake itself. Previous attempts along the "defensive stop-before-start" and "identity-based singleton cleanup" lines (fix: fix SubprocessASGIServer singleton cleanup causing test_prefect_test_harness flakiness #21414, fix: fix flaky tests on main (harness singleton, SQLite lock, timing tolerance) #21460) were closed without merging, and I don't yet understand the root cause well enough to propose a different fix. That work belongs in a separate investigation.

Relationship to prior PRs

• Builds on Fix flaky tests: SubprocessASGIServer singleton cleanup and SQLite lock retry #21405 (merged). Fix flaky tests: SubprocessASGIServer singleton cleanup and SQLite lock retry #21405's `_instance_key` cleanup only runs if `stop()` actually runs. This PR makes `stop()` actually run on exception paths.
• Does not overlap with fix: properly drain workers in prefect_test_harness when used in async contexts #19770 (async drain refactor). The exact `drain_workers` coroutine and `run_coro_as_sync` wrapper from fix: properly drain workers in prefect_test_harness when used in async contexts #19770 are preserved verbatim, just moved onto the ExitStack.
• Does not overlap with Fix EventsWorker singleton leak in prefect_test_harness #19343 (EventsWorker singleton leak). `EventsWorker.drain_all()` still happens in the drain callback.

Tests

One new regression test, verified to fail against the pre-fix source:

• `test_prefect_test_harness_cleans_up_on_exception_in_body` — raises from inside `with prefect_test_harness():`, asserts `SubprocessASGIServer._instances` has no `None` key after the unwind, then proves a subsequent harness actually works.

Test plan

• `tests/testing/test_utilites.py` — 9/9 pass with the fix
• Regression test verified to fail on pre-fix source
• Ruff + format clean

[codspeed-hq]

Merging this PR will not alter performance

✅ 2 untouched benchmarks

---

_{Comparing fix/prefect-test-harness-exception-cleanup (6d5befb) with main (f6addff)}