feat(training): improve self-play progress visibility and reward diagnostics

Update training config, add checkpointing on HF

test hf space commit

Sync current main to Hugging Face Space

fix(rewards): never crash GRPO on malformed completions