feat(embedding): add WeMM-Embedding support

[Minamiyama]

Support multimodal text, image, video, and interleaved inputs through sentence-transformers and vLLM with Hugging Face and ModelScope sources.

Summary

Add built-in support for Tencent's multimodal WeMM-Embedding model family:

• WeMM-Embedding-2B
• WeMM-Embedding-4B
• WeMM-Embedding-9B

Models are available from:

• Hugging Face, revision main
• ModelScope, revision master

Supported embedding engines are intentionally limited to existing Xinference backends:

• sentence_transformers
• vllm

Model Metadata

Model	Dimensions	Max tokens
WeMM-Embedding-2B	2048	262144
WeMM-Embedding-4B	2560	262144
WeMM-Embedding-9B	4096	262144

[gemini-code-assist]

Code Review

This pull request adds support for the WeMM-Embedding model family (2B, 4B, and 9B) in Xinference, enabling multimodal inputs (text, image, and video) across both the sentence_transformers and vllm engines. It updates API schemas, restful clients, and core embedding logic to support structured multimodal inputs, and includes a PyAV-based video reader fallback. The review feedback correctly points out that the specified dependency versions for sentence-transformers (>=5.7.0) and transformers (>=5.2.0) do not exist on PyPI and must be updated to valid, existing versions.

[qinxuye]

LGTM