[model] Add Gemma 4 support

[marksverdhei]

Summary

• Ports upstream PR [model] gemma4 hiyouga/LlamaFactory#10346 to add Gemma 4 model family support
• Adds Gemma4 and Gemma4n multimodal plugins (image processing with pan-and-scan, audio support for Gemma4n)
• Adds chat templates, tool calling utilities, model constants, and visual module support
• Includes HyperParallel workflow compatibility fix

Upstream status

The upstream PR is still open/WIP with known limitations:

• gemma4-e*B variants are not yet ready for training
• Gemma4n audio plugin and forwarding tests are pending

Files changed

File	Change
data/mm_plugin.py	Gemma4/Gemma4n multimodal plugins
data/template.py	Chat templates for gemma4, gemma4n
data/tool_utils.py	Tool calling support
extras/constants.py	Model registry entries
model/model_utils/visual.py	Visual module support
train/hyper_parallel/workflow.py	HyperParallel compatibility
tests/data/test_mm_plugin.py	Plugin tests

Test plan

• Verify Gemma 4 text-only SFT training works
• Test Gemma 4 multimodal (image) training
• Test Gemma 4n with audio inputs
• Verify no regressions on existing models