Improve out-of-the-box recipe for Gemma4

[faradawn]

Recipe file: Google/Gemma4.md

Goal: 1) no arbitrary flags and incorrect parameters, 2) chat templates and parser names are up-to-date with the vllm codebase, 3) deprecated fields are replaced with current equivalents.

Action items

• Outdated chat template and deprecated reasoning_content field: The recipe used an old chat template and the deprecated reasoning_content field for reasoning/tool calling. Updated to the new chat template merged in [Tool] adjust_request to reasoning parser, and Gemma4 fixes vllm#39027 and replaced with the current reasoning field. Solved by vllm-project/recipes#322 (@bbrowning).
• Typo in --limit-mm-per-prompt flag: The multimodal prompt limit flag used a space instead of a comma to separate modalities, causing a parse error: Value image=0 cannot be converted. Solved by vllm-project/recipes#310 (@GaurangTandon).
• Incorrect reasoning and tool parser names: The recipe specified wrong parser names for the Gemma4 reasoning and tool parsers, inconsistent with the actual parser code. Solved by vllm-project/recipes#316 (@whynotkimhari).

[faradawn]

Docs have been compiled and tested with

  mkdocs serve                                                                                                                                                                                                                                                                                     

[FredericOdermatt]

The template is out of sync again (due to HF side updates), I've allowed myself to open a PR here vllm-project/vllm#39570