support gemma4 vllm multi-modal inference

[hjh0119]

No description provided.

[gemini-code-assist]

Code Review

This pull request introduces support for vllm mode within the Gemma4Template, including environment-based configuration for video and image processors and specialized tag replacement logic for multimodal inputs. It also updates the load_audio utility to handle string paths more effectively. A critical issue was identified in the replace_tag method where load_audio is called unconditionally; this will lead to a crash in vllm mode because the audio data is already pre-processed into a format incompatible with librosa.load.