fix(rollout): clear stale LLM trajectory on rollout reuse

[zhiheng-yang]

Problem

_init_rollout() can be called with explicit job and rollout names, so the target directory may already exist.
In that case, an old trajectory/llm_trajectory.jsonl is left in place.

Usually live capture replaces the file after the first completed model call.
If the new run fails before then, nothing overwrites it.
Training export and artifact inspection can then read the previous run's trajectory as if it came from the current attempt.

I reproduced this by initializing the same rollout path twice and placing a trajectory in between.
After the second initialization:

• main: stale_exists=True
• this patch: stale_exists=False

Change

Remove this one attempt-owned file when the rollout directory is initialized.
The rest of the directory is left alone, and the separate continue-run flow is unchanged.

• src/benchflow/rollout/_setup.py removes the stale trajectory.
• tests/test_sdk_internals.py covers reuse of the same rollout path.

Tests

uv run pytest -q tests/test_sdk_internals.py

Result: 51 passed.

[devin-ai-integration]

🔍 Devin Review: 1 flag

Not posted on this PR by your GitHub settings — view it in Devin Review. (Configure)