Reported by @smallhadroncollider on v0.6.49: passing
``chat_template_kwargs: {enable_thinking: false}`` to
/v1/chat/completions was silently ignored, so qwen3.6-27b-8bit burned
~3000 reasoning tokens for a one-line joke.

Root cause: ``ChatCompletionRequest`` did not declare
``chat_template_kwargs`` at all. Pydantic accepted the payload but
dropped the unknown field on parse, so the chat template ran with
thinking enabled by default. llama.cpp + vLLM upstream both honor
this OpenAI extended-spec field.

Fix:
- Add ``chat_template_kwargs: dict | None = None`` to
  ``ChatCompletionRequest`` (api/models.py)
- Add ``_extract_thinking_from_request(request)`` — request-only
  precedence (ctk.enable_thinking → request.enable_thinking → None),
  with string-bool tolerance ("true"/"false") for client-friendliness.
  Single source of truth, used by both wrappers below.
- Add ``_resolve_enable_thinking(request)`` — adds the cfg.no_thinking
  layer on top of the extractor for the OpenAI/anthropic routes.
- Wire the helper into routes/chat.py and routes/anthropic.py (×2 call
  sites — non-streaming and streaming).
- routes/chat.py: also resolve the max_tokens default through the
  same thinking flag (was based on top-level enable_thinking only,
  giving the smaller thinking-off budget when ctk said thinking-on).
- speculative/dflash/server.py: layers its own closure ``no_thinking``
  arg on top of ``_extract_thinking_from_request`` (NOT the cfg-aware
  helper — dflash never sets cfg.no_thinking, and consulting it would
  inject a hidden global-state dependency).

Precedence (OpenAI / anthropic routes):
  1. server --no-thinking (cfg.no_thinking) → False
  2. request.chat_template_kwargs["enable_thinking"]
  3. request.enable_thinking (top-level field, our extension)
  4. None (template default)

Precedence (dflash route):
  1. dflash closure --no-thinking → False
  2. request.chat_template_kwargs["enable_thinking"]
  3. request.enable_thinking
  4. None

Verification:
- ruff check + format clean
- 23 new tests in tests/test_chat_template_kwargs.py covering: request
  model field accept, OpenAI/anthropic precedence (10 cases incl.
  string-bool tolerance + garbage-value handling), the shared extractor
  (6 cases), dflash inline precedence (4 cases incl. the regression
  guard for "dflash must NOT consult cfg.no_thinking even if set")
- pytest tests/test_chat_template_kwargs.py: 23 passed
- Full unit suite: 3474 passed (3451 baseline + 23 new), 19 skipped,
  7 deselected, 1 xfailed — no regressions vs main
- DeepSeek V4 Pro adversarial review: 3 rounds to convergence
  (round 1 → dflash hidden cfg dependency, round 2 → drift risk +
  missing dflash test, round 3 → "No blocking issues found.")

Closes #387.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>