feat(embedding): add EmbeddingGemma 2 support

[Minamiyama]

Description:

Add google/embeddinggemma-2 as a built-in model with Hugging Face (main) and ModelScope (master) download sources.

• Add SentenceTransformers, Transformers, and vLLM adapters for text, image, audio, video, and mixed inputs.
• Support task prompts, an 8192-token context, and 128/256/512/768-dimensional embeddings with normalization after truncation.
• Use BF16/FP32 and reject FP16, which produces invalid embeddings.
• Preserve system Torch/TorchVision pins for native backends to prevent incompatible virtualenv dependencies.
• Provisionally target vllm>=0.32.0 and verify that EmbeddingGemma2Model is registered.

[qinxuye]

LGTM. Verified the native adapters, multimodal input routing, pooling, dimensions, and dtype handling. All 24 EmbeddingGemma 2 tests passed locally, including the native Transformers/SentenceTransformers mixed-input test. The Windows 3.10 CI timeout is in the unchanged SGLang/Xavier P/D tests; vLLM GPU inference was not run locally.