Bound full GC passes in tests to one test's allocations

[DABH]

What was changed

• tests/conftest.py: a pytest_collection_finish hook runs gc.collect() then gc.freeze() once, after all test modules are imported.
• tests/contrib/langgraph/test_replay.py: asserts the replay imports nothing into the sandbox after initial workflow load (same guard as the OpenAI Agents replay test).

Why

test_replay tripped the deadlock detector on the 3.10 macOS runner while the first activation was in StateGraph.compile(); the interrupt landed in a weakref callback inside ast.parse, i.e. during a cyclic GC pass. That activation does 3-8 ms of work and imports nothing into the sandbox. The test process heap grows to ~1.7M objects, and a full GC pass over it (160-290 ms locally, several times that on a 3-core runner shared by three xdist workers) runs on whichever thread allocates, so first activations occasionally absorb it inside the 2 s budget. Freezing the import-time heap once removes it from every later full pass. An earlier revision froze after every test; that made each test's survivors permanent (RSS 77 MB to 3.6 GB over 229 serial tests, versus a 490 MB plateau without the hook) and killed the 2-core ubuntu-arm CI runners mid-run on every attempt.

Testing

Serial tests/worker/test_workflow.py with RSS sampled: no hook plateaus at 490 MB, single freeze at 390 MB, per-test freeze climbs to 3.6 GB. Full gc.collect() per test on a 51-test subset: 70 ms mean / 179 ms max without the hook, 37 ms / 122 ms with the single freeze. LangGraph replay and sandbox suites pass on 3.10 and 3.14; poe lint clean.

[DABH]

Reworked in 764aa2b (see updated description): a single post-collection freeze replaces the per-test freeze, which leaked each test's survivors and exhausted the 2-core ubuntu-arm runners.