Harden SGLang decode training sample handling

[mesaleh]

Summary

• add decode min_new_tokens and stop_token_ids config plumbing for SGLang decode-mode training
• skip prompts that cannot leave room for the minimum generated tokens within max_seq_length
• drop zero-loss completions and completions that begin with a configured stop token while preserving batch positions
• make HFRunner an optional import so SGLang decode unit tests do not require Mooncake at import time

Testing

• .venv/bin/python -m pytest tests/test_sgl_engine_decode.py -q