fix: cancel pending rollout tasks after a failure

[yuyz-cyber]

Problem

Training and evaluation launch rollout tasks concurrently at three levels: samples within a rollout group, rollout groups within a training batch, and samples within an evaluation batch.

When one task raises an exception, the exception is propagated but the unfinished sibling tasks are not cancelled. They can continue generation or reward computation after the rollout has failed, wasting resources and potentially leaving unobserved task exceptions in the event loop.

Fix

At each concurrency boundary, cancel unfinished sibling tasks and wait for their cleanup before re-raising the original exception. The training path also resets its shared generation state, and both training and evaluation close their progress bars before returning the failure. Successful rollout behavior is unchanged.

Test plan

• Added a regression test where one generation task fails while a sibling is still running.
• Verified that the sibling is cancelled and the original exception is preserved.
• tests/plugin_contracts/test_plugin_generate_contracts.py: 5 passed.
• tests/test_streaming_rollout.py: 26 passed.
• Ruff, Black, and isort checks passed.