Bug fixes for issues #48-#65:
- render_dummy_text cycles tokens for any max_tokens value
- Streaming requests logged to JSONL
- max_model_len enforced in all code paths
- prefill/app.py: /ping and /metrics endpoints
- Warmup penalty works with scheduling
- /v1/completions routes through scheduler
- Scheduler sync-safe (no event loop dependency)
- CLI arg priority fixed (_TrackAction subparser bug)
- Batch metrics always present in /metrics
- /debug/batch returns zero-state when scheduling disabled
- Prefill batch FIFO ordering fixed
- PyPI publish workflow added
