feat: add Gemma 4 tool call and reasoning parsers

[mikepixelmagic-dev]

Summary

Adds native Gemma 4 support for both tool calling and reasoning extraction.

• Tool call parser (--tool-call-parser gemma4): Parses Gemma 4's <|tool_call>call:func{key:<|"|>value<|"|>}<tool_call|> format. Supports hyphenated/dotted function names (e.g. searxng-search).
• Reasoning parser (--reasoning-parser gemma4): Extracts <|channel>thought...<channel|> tags into the reasoning_content API field. Stateful streaming parser handles tags that span multiple MLX tokens.

Usage

vllm-mlx serve mlx-community/gemma-4-26b-a4b-it-4bit \
  --enable-auto-tool-choice \
  --tool-call-parser gemma4 \
  --reasoning-parser gemma4

Files changed

File	What
vllm_mlx/tool_parsers/gemma4_tool_parser.py	Tool call parser (new)
vllm_mlx/tool_parsers/__init__.py	Register import
vllm_mlx/reasoning/gemma4_parser.py	Reasoning parser (new)
vllm_mlx/reasoning/__init__.py	Register in builtin parsers

Test plan

• Non-streaming tool calling with standard function names
• Non-streaming tool calling with hyphenated names (e.g. searxng-search)
• Streaming tool call extraction
• Non-streaming reasoning extraction (reasoning_content populated, content clean)
• Streaming reasoning extraction (no <|channel> tag leakage)
• Tested on Apple M5 Max with mlx-community/gemma-4-26b-a4b-it-4bit

Fixes #264

🤖 Generated with Claude Code

[waybarrios]

The reasoning parser landed via #268 and the tool call parser is covered by #269. Closing in favor of those two. Thanks for putting this together.