fix: cancel in-flight model task when a parallel input guardrail errors

[bymle]

Summary

When input guardrails run in parallel with the model turn (run_in_parallel=True), the model turn is started with asyncio.create_task(...) and awaited together with the guardrails via asyncio.gather(...) in AgentRunner.run (src/agents/run.py).

The existing handler only catches InputGuardrailTripwireTriggered, in which case it cancels and drains the in-flight model_task:

except InputGuardrailTripwireTriggered:
    if should_cancel_parallel_model_task_on_input_guardrail_trip():
        if not model_task.done():
            model_task.cancel()
        await asyncio.gather(model_task, return_exceptions=True)
    ...
    raise

If a parallel guardrail (or the model turn) raises any other exception — e.g. a bug in a user's guardrail function, a ValueError, a transient error — asyncio.gather propagates that exception but does not cancel the sibling awaitables, so model_task is left running after run() has already raised. That orphaned task keeps a live model request going and surfaces as Task was destroyed but it is pending! / Task exception was never retrieved warnings, plus wasted tokens/quota.

This adds a generic except clause that performs the same flag-gated cancel-and-drain of the model task as the tripwire path, so a non-tripwire guardrail error no longer leaks the in-flight model task. The behavior on the tripwire path is unchanged, and the should_cancel_parallel_model_task_on_input_guardrail_trip() gate (Temporal replay compatibility) is respected in both paths.

Test plan

• Added test_parallel_guardrail_non_tripwire_error_cancels_model_task in tests/test_guardrails.py, mirroring the existing test_parallel_guardrail_trip_cancels_model_task but with a parallel guardrail that raises a non-tripwire RuntimeError after the model starts. It asserts the error propagates from Runner.run and that the in-flight model task was cancelled.
• Verified the test fails on main (assert model_cancelled.is_set() is True → assert False is True, i.e. the model task was orphaned) and passes with this change.
• make format, make lint, make typecheck clean; full suite passes (4625 passed, 5 skipped).

Issue number

N/A (found via code review).

Checks

• I've added new tests (if relevant)
• I've added/updated the relevant documentation (no user-facing docs affected)
• I've run make lint and make format
• I've made sure tests pass