Add CustomVoice, VoiceDesign, native clone, and fixed-mode HTTP server support

[AntNoHuabei]

Summary

This PR expands Qwen3-TTS runtime coverage and packaging around three practical synthesis modes:

• CustomVoice: built-in speaker IDs / speaker names with optional style instruction
• VoiceDesign: instruction-driven voice design models
• VoiceClone: Python-free WAV / prompt based voice cloning

It also adds a fixed-mode HTTP server binary (qts_server) with OpenAI-compatible speech output and async job progress APIs.

Major changes

Voice modes

• Add CustomVoice model metadata loading and speaker-token conditioning.
• Add VoiceDesign synthesis APIs and CLI support.
• Add native WAV voice-clone prompt construction without requiring Python at runtime.
• Support both x-vector-only clone and ICL clone with reference text/audio codes.
• Add support for reusable protobuf voice-clone prompt files.

Runtime and model compatibility

• Add VoiceDesign model kind detection.
• Support VoiceDesign GGUF variants that use code_pred.hidden_proj.* naming.
• Add Q8_0 tensor dequantization for F32 loading paths used by VoiceDesign/custom projection paths.
• Add vocoder artifact fallback discovery for sibling model directories.
• Add native soxr-based resampling and audio-code extraction path.

Performance and Vulkan stability

• Improve CustomVoice Vulkan talker stability.
• Optimize CustomVoice code predictor path.
• Add optional preprojected code predictor embeddings.
• Fuse CustomVoice code predictor projections where possible.
• Keep CPU fallback behavior for paths that are not safe or profitable on Vulkan.

Server mode

• Add standalone qts_server.exe binary.
• Server mode is fixed at startup: none, custom, design, or clone.
• Add OpenAI-compatible synchronous endpoint:
  • POST /v1/audio/speech
• Add QTS async job APIs:
  • POST /v1/qts/audio/jobs
  • GET /v1/qts/audio/jobs/{job_id}
  • GET /v1/qts/audio/jobs/{job_id}/audio
• Add synthesis progress reporting with rollout frame progress.
• Serialize requests through a single worker so the loaded ggml/Vulkan engine is not concurrently accessed.

CLI / server examples

qts_server --mode design \
  --model-dir models/Qwen3-TTS-12Hz-1.7B-VoiceDesign \
  --backend vulkan \
  --vocoder-ep cpu \
  --language-id 2055

qts_server --mode custom \
  --speaker serena \
  --model-dir models/Qwen3-TTS-12Hz-1.7B-CustomVoice

qts_server --mode clone \
  --voice-clone-wav ref.wav \
  --voice-clone-ref-text "reference text" \
  --model-dir models/Qwen3-TTS-12Hz-0.6B-Base

Validation

Tested locally on Windows with Vulkan:

• CustomVoice synthesis
• VoiceDesign synthesis
• WAV clone / ICL clone path
• qts_server fixed design mode
• /health
• /v1/audio/speech
• async job creation, progress polling, and WAV download
• fixed-mode rejection when a request tries to switch mode

Build/check commands used:

cargo check -p qts_cli --no-default-features
cargo build --release -p qts_cli --bin qts_server --no-default-features --features vulkan

Notes

• qts_server intentionally fixes conditioning mode at startup to keep server behavior predictable for wrappers and API gateways.
• The OpenAI-compatible endpoint remains synchronous; progress is exposed through the QTS async job APIs.
• Runtime packaging for Windows server currently needs only:
  • qts_server.exe
  • onnxruntime.dll
  • soxr.dll