feat: add Gemma 4 tool call parser

[keegoid]

Summary

• Adds a new gemma4_tool_parser.py that handles Gemma 4's unique tool calling format (<|tool_call>call:name{key:<|"|>value<|"|>}<tool_call|>)
• Implements a state-machine converter for Gemma's custom argument syntax (unquoted keys, <|"|> string delimiters) to valid JSON
• Integrates Gemma 4 detection into the auto parser (both streaming and non-streaming paths)
• Adds gemma4 to CLI --tool-call-parser choices
• 21 tests covering arg conversion, nested objects, arrays, hyphenated tool names, streaming, think tag stripping, and auto-detection

Background

Gemma 4 models use a non-standard tool call format defined in their chat template. Unlike other model families that use JSON or XML, Gemma 4 uses special tokens (<|tool_call>, <tool_call|>, <|"|>) for structure and string quoting. This parser converts that format to OpenAI-compatible tool_calls responses.

Tested end-to-end with mlx-community/gemma-4-31b-it-8bit on Apple Silicon (M5 Max 128GB):

• Single and multi-tool calls
• Multi-turn with tool responses
• Vision + tool calling combined
• Streaming

Usage

vllm-mlx serve mlx-community/gemma-4-31b-it-8bit --port 8000 --mllm   --enable-auto-tool-choice --tool-call-parser gemma4

Test plan

• All 21 new Gemma 4 parser tests pass
• All 78 existing tool parser tests pass (99 total, 0 failures)
• End-to-end tool calling verified against live Gemma 4 31B model
• Vision + tool calling combo verified
• Streaming tool call buffering verified
• Gemma 4 E4B model (not yet tested, same format expected)

🤖 Generated with Claude Code

[waybarrios]

The auto-detection logic you built for AutoToolParser was the right call. I pulled that approach into #269 so Gemma 4 tool calls work out of the box with --enable-auto-tool-choice. The standalone parser is also covered there. Thanks for flagging the RotatingKVCache issue on #256 too, that was a key catch. Closing in favor of #269.