[Fix] Fix SGLang tensor parallel rank bug and improve code quality

[yurekami]

Summary

This PR fixes a critical bug in SGLang's tensor parallel support and improves overall code quality.

Bug Fix

• Critical: Fix start_worker_listener_thread() call to use tp_rank instead of torch.cuda.current_device() in SGLang integration. The socket path is based on rank, not device ID, so using the wrong value causes IPC communication failures in multi-GPU tensor parallel setups.
• Add is_worker parameter to SGLang's init_kvcached() to match vLLM's behavior - only workers should start the listener thread.
• Add _tp_size global state to SGLang interfaces and pass it to KVCacheManager for proper tensor parallel coordination.
• Update SGLang patches to detect tp_rank and tp_size from SGLang's distributed state via get_tp_group().

Code Quality Improvements

• Replace bare except clauses in autopatch.py with specific exception handling (ImportError vs other exceptions) and add debug/warning logging.
• Add comprehensive docstrings to all public API functions in both vLLM and SGLang integration modules.
• Add type hints to global module state variables.
• Add input validation for limit-percent CLI command (0-100 range).

Test plan

• Verify Python syntax is valid (checked locally with py_compile)
• Test with SGLang single-GPU inference
• Test with SGLang tensor parallel (multi-GPU) inference
• Test with vLLM to ensure no regressions

🤖 Generated with Claude Code