bugfix: fix OpenAI multi-turn tool calls and Content-Length FATAL crash.

[shifengmin]

Summary

• Preserve tool_calls / reasoning_content / tool_call_id in OpenAI chat messages, parse tool-call arguments as JSON objects for Jinja templates, and strip leaked GLM control tokens from streamed output.
• Treat missing/invalid Content-Length as a request error (SetFailed) instead of LOG(FATAL), and use application/json for non-stream HTTP responses.
• Bump third_party/minja for GLM-5.2 Jinja syntax support (.0 member access / capitalize filter).

Test plan

• Multi-turn tool-call chat against GLM returns correct history (no dangling <|observation|> / leaked control tokens)
• POST /v1/chat/completions without Content-Length returns error and process stays up
• Normal chat with Content-Length still succeeds
• Build xllm_master_serving on NPU/dev host

Made with Cursor