feat: add Gemma 4 tool call parser

[jackneil]

Summary

• Adds gemma4_tool_parser.py — parses Gemma 4's <|tool_call>call:name{<|"|>key<|"|>:val}<tool_call|> format into OpenAI-compatible tool_calls
• Handles nested objects, arrays, unicode, braces in strings, quoted keys, multiple calls per block, think tags
• Registers as --tool-call-parser gemma4
• Adds streaming filter tags for <|tool_call>/<tool_call|>
• 30 tests including OpenAI format integration tests verifying exact Claude Code compatibility

Usage

vllm-mlx serve google/gemma-4-27b-it \
  --enable-auto-tool-choice --tool-call-parser gemma4

Test plan

• 30 unit/integration tests passing
• OpenAI format verified (content=null, finish_reason=tool_calls, function.arguments as JSON string)
• End-to-end with Claude Code

🤖 Generated with Claude Code

[waybarrios]

@jackneil I rebased this on main (clean, no conflicts) and added one commit on top: Gemma 4 auto-detection in AutoToolParser, based on keegoid's approach from #254. This way Gemma 4 tool calls work out of the box with --enable-auto-tool-choice without needing --tool-call-parser gemma4.

What I added:

• Gemma 4 as the first format tried in AutoToolParser.extract_tool_calls()
• <|tool_call> and <tool_call|> in the streaming markers/end_markers
• Imports Gemma4ToolParser and delegates to it when the marker is detected

All 121 tool parser tests pass, black and ruff clean.

[waybarrios]

Rebased, added auto-detection from #254, 121 tests pass. Ready to merge.

[waybarrios]

@Thump604 would you mind taking a look at this one? It's the Gemma 4 tool call parser from Jack, plus I added auto-detection in AutoToolParser based on keegoid's approach from #254. This is the last piece to complete the Gemma 4 support alongside #268 and #256 that already landed.