feat: migrate Qwen3 reasoning parser to token-level OutputRouter (closes #64)

[Ai-chan-0411]

Closes #64

Summary

Migrate Qwen3's <think>/</think> reasoning parser from regex-based BaseThinkingReasoningParser to the token-level OutputRouter state machine.

Changes

vllm_mlx/output_router.py

• Add <think>/</think> token handling in feed() — think_start enters THINKING state, think_end switches to CONTENT
• Add Qwen3/DeepSeek auto-detection in from_tokenizer() via <think> + </think> vocabulary entries
• Token map already had think_start/think_end fields (added for future migration) — now wired up

tests/test_output_router.py

• Add Qwen3 vocabulary fixtures and qwen3_router fixture
• Add TestQwen3ThinkRouting class with 9 tests:
  • <think> enters THINKING state
  • </think> switches to CONTENT state
  • Tokens between tags routed to REASONING channel
  • Content tokens after </think> routed to CONTENT channel
  • Full <think>reasoning</think>content sequence via feed_sequence()
  • Implicit thinking (no <think>, only </think>) handled
  • No-tag output passes through as pure content
  • Control tokens (bos/eos) suppressed
  • Reset clears thinking state
• Add test_qwen3_detected in TestFromTokenizer

31 tests total, all passing.

Design

Follows the same pattern as the existing Gemma 4 implementation — token-level state machine with no text-level regex matching. This eliminates partial-token split issues that affect the current regex-based parser.

The detection uses <think> + </think> in the tokenizer vocabulary, which covers both Qwen3 and DeepSeek R1 model families.

[Ai-chan-0411]

Closing as maintainer has not reviewed after 168h. Thank you for the opportunity!