fix(fast-llm): stop producers and helpers on training completion

[jlamypoirier]

gpt-5 Codex note:

This is a fixup for PR #142. It extends the Fast-LLM early-stop fix beyond the actor path.

Changes:

• The preprocessor now uses the explicit Fast-LLM training_finished event before stopping, instead of the legacy sample-counting heuristic.
• The launcher now stops remaining supervised helper processes after the finetune process has exited cleanly and training completion has been observed. This covers redis/actor/preprocessor-style helpers, not only inference servers.

Why:

• In Fast-LLM mode, samples_processed reflects Redis entries read, not natural optimizer completion, so preprocessing could still stop converting actor rollouts into fast_llm_streaming documents too early.
• After a clean Fast-LLM completion, non-inference helper processes can otherwise keep the launcher alive even though the trainer has finished.

Verification from the local branch:

• /Users/joel.lamy-poirier/Projects/Fast-LLM/venv/bin/python -m py_compile pipelinerl/launch.py pipelinerl/preprocess.py tests/test_launch_process_monitoring.py
• git diff --check
• Targeted pytest collection is blocked in this local environment because tests/conftest.py imports pipelinerl.vllm1, which requires uvloop.