feat: [new model] minimax music3

[Minamiyama]

Summary

This PR adds initial MiniMax-Music3 support as a built-in audio model with the new text2music ability.

• Adds MiniMax-Music3 using Diffusers ModularPipeline
• Supports Hugging Face and ModelScope model sources
• Adds a dedicated Music Generation Web UI
• Reuses the OpenAI-compatible POST /v1/audio/speech endpoint
• Adds user documentation and request examples

Backend

• Model family: minimax_music3
• Model ability: text2music
• Engine: Diffusers only
• Accelerator: NVIDIA CUDA only
• Default dtype: BF16
• Supports optional CPU and group offload
• Loads pipeline components from the local Xinference model cache to avoid duplicate downloads
• Pins the Diffusers commit containing MiniMax-Music3 ModularPipeline support
• Preserves the pipeline's native 44.1 kHz, stereo, float32 audio without resampling or quantization
• Encodes the result as a standards-compliant IEEE-float WAV

The model is registered under the MiniMax-Music3 Community License.

API

Music generation reuses /v1/audio/speech:

• input: lyrics
• kwargs.instruct: music description
• kwargs.seed: non-negative random seed, default 0
• kwargs.duration: maximum duration in seconds, default 60, range 0.04 to 360
• response_format: wav only

Unsupported speech or sampling parameters are rejected with model-specific validation errors.

Web UI

Adds a distinct Music Generation capability with:

• Multiline lyrics input
• Music description input
• Seed input with a random-seed button
• Duration in seconds
• WAV audio playback

Regular TTS and voice-cloning forms retain their existing behavior.

[gemini-code-assist]

Code Review

This pull request integrates the MiniMax-Music3 text-to-music model into Xinference, adding backend support via the Diffusers ModularPipeline engine, updating model specifications, and introducing a new 'text2music' capability. It also updates the frontend UI to support music generation with localized translations and dedicated form fields. The review feedback suggests fixing a conflict between group offloading and CPU offloading in the backend, supporting 'bf16' as a data type alias, using safer number parsing with fallbacks in the frontend configuration, improving the lyrics input placeholder text, and adding the new model to the legacy dispatch chain for robustness.

[qinxuye]

LGTM