Forward chat template kwargs in batched chat

[krystophny]

Summary

• accept chat_template_kwargs in the OpenAI chat completion request model
• forward those kwargs through the server into the batched engine chat-template path
• cover the request-model, engine, and endpoint forwarding behavior with regressions

Why

Tabura relies on chat_template_kwargs.enable_thinking=false for Qwen 3.5 local turns. In continuous batching mode the request field was dropped before it reached the tokenizer, so Qwen kept emitting thinking text even when the client explicitly disabled it.

Verification

Test fails on main

$ cd /tmp/vllm-mlx-main.dV8sAj
$ PYTHONPATH=/tmp/vllm-mlx-main.dV8sAj "$HOME/Library/Application Support/tabura/llm/venv/bin/python" -m pytest /tmp/vllm_mlx_chat_template_regression.py -q
F                                                                        [100%]
=================================== FAILURES ===================================
...
E       KeyError: 'chat_template_kwargs'
...
1 failed, 3 warnings in 2.46s

Test passes after fix

$ cd /tmp/vllm-mlx-fork
$ PYTHONPATH=/tmp/vllm-mlx-fork "$HOME/Library/Application Support/tabura/llm/venv/bin/python" -m pytest /tmp/vllm_mlx_chat_template_regression.py -q
.                                                                        [100%]
1 passed, 3 warnings in 2.43s

$ PYTHONPATH=/tmp/vllm-mlx-fork "$HOME/Library/Application Support/tabura/llm/venv/bin/python" -m pytest tests/test_chat_template_kwargs.py -q
...                                                                      [100%]
3 passed in 2.10s