…thropic-style thinking field

vLLM is an OpenAI-compatible server, so it speaks OpenAI's chat
completions protocol and does not understand Anthropic-native extension
fields. The Vllm branch of apply_reasoning_effort was previously injecting
`thinking: {type: "disabled" | "enabled"}` at the top level of the request
body — but vLLM silently ignores this unknown field.

For Qwen3 (and other reasoning-capable models hosted via vLLM), the
canonical way to toggle thinking is the OpenAI-protocol extension
`chat_template_kwargs.enable_thinking`, which vLLM forwards into the
model's chat template. See the vLLM docs on chat templating + Qwen3
official template's `enable_thinking` parameter.

Symptom before this fix
-----------------------
With reasoning_effort="off" against a Qwen3 deployment, the model still
generated 10+ seconds of reasoning tokens, which vLLM placed into the
non-OpenAI-standard `reasoning` field. This client does not consume
`reasoning`, so the user observes a ~13s "freeze" before any content
delta arrives. Manual curl reproduction:

    $ curl http://vllm-host:8000/v1/chat/completions -d '{
        "model": "...", "messages":[{"role":"user","content":"hi"}],
        "max_tokens": 100
      }'
    -> reasoning: "Thinking Process: 1. Analyze the User's Input..."
    -> content: null   (max_tokens spent in reasoning)

    $ curl http://vllm-host:8000/v1/chat/completions -d '{
        "model": "...", "messages":[{"role":"user","content":"hi"}],
        "max_tokens": 100,
        "chat_template_kwargs": {"enable_thinking": false}
      }'
    -> reasoning: null
    -> content: "你好！我是 Qwen…"   (full answer, < 5s)

After this fix
--------------
Vllm branch now mirrors NvidiaNim (which already uses
chat_template_kwargs, but with the NVIDIA-specific `thinking` key rather
than the Qwen-standard `enable_thinking`).

End-to-end measurement against vLLM hosting Qwen3.6-35B-A3B-FP8:
- TTFT (time-to-first-text-delta): 13039ms -> 274ms
- Total LLM call: 13s -> 5.7s
- Effective output rate: 3 chars/s (most time in hidden reasoning)
                          -> 46 chars/s

Note: Sglang and other OpenAI-compatible servers (Fireworks, Novita)
likely have the same issue but I have not verified the right field for
each. They are left unchanged in this PR; the Vllm branch is the
minimal, verified fix.