training: summarize HF job event logs

[AbdelStark]

Summary

• Add codelewm.training.job_events to parse live CODELEWM_JOB_EVENT log lines and persisted reports/job_progress.jsonl rows.
• Add scripts/hf-job-event-status for compact one-shot or watched status summaries across HF job IDs, including latest progress, losses, ETA, CUDA memory, checkpoint state, and the no-collapse gate.
• Document the status command in the HF operation and scaled-training runbooks, with docs tests enforcing the helper.

Live check

The new helper was exercised against the replacement v0.8 A10G jobs from #370:

uv run scripts/hf-job-event-status 6a2278d2e6aa50b87b9eba56 6a227a6ce52fdd2a02ed9005

At the check, both jobs were RUNNING, emitting progress, and had passing step-1000 collapse diagnostics.

Validation

• uv run pytest tests/training/test_job_events.py tests/docs/test_scaled_training_runbook.py tests/docs/test_hf_ml_intern_training.py
• uv run pytest tests/training/test_job_events.py tests/training/test_execution_runner.py tests/training/test_execution_train_config.py tests/docs/test_scaled_training_runbook.py tests/docs/test_hf_ml_intern_training.py
• uv run pytest tests/training/test_job_events.py
• uv run python -m compileall -q codelewm/training/job_events.py scripts/hf-job-event-status tests/training/test_job_events.py
• git diff --check
• uv run codelewm secret-scan codelewm/training/job_events.py scripts/hf-job-event-status tests/training/test_job_events.py docs/training/SCALED_TRAINING_RUNBOOK.md docs/operations/HF_ML_INTERN_TRAINING.md --json
• direct token-pattern grep over the touched Python/script/test files found no matches

Part of #370; supports #364.