Add native Higgs Audio v3 TTS support

[Saganaki22]

Summary

Adds native higgs_tts support for Higgs Audio v3 TTS 4B, including text-to-speech and zero-shot voice cloning with --voice-ref.

Changes

• Adds Higgs model assets, tokenizer, generator, codec, loader, and session runtime.
• Registers higgs_tts in the default model registry and CMake build.
• Adds higgs_audio_v3_tts_4b to tools/model_manager.py.
• Adds optional Qwen3 ASR guidance for users who need reference transcripts.
• Adds tools/prepare_voice_ref.py for normalizing WAV/MP3/M4A/FLAC reference audio to mono 24 kHz WAV.
• Documents Higgs usage in docs/tts.md.

Testing

• Built audiocpp_cli with CUDA on Windows.
• Verified model-manager info for:
  • higgs_audio_v3_tts_4b
  • qwen3_asr_0_6b
• Ran local Higgs TTS voice generation & voice clone tests with reference audio.
• Verified generated speech content locally with Whisper.

compressO-higgs_cuda_10s_take3.mp4

[0xShug0]

@Saganaki22 Thank you for the effort, but I already have this implementation ready. In general, any model that already has a folder under tests/ has already been implemented.

Also please check #19 (comment) for the metrics I am looking for. If your impl is faster enough and memory stable, I will merge.

[0xShug0]

I realized I pushed an old version of the Higgs Audio test to `tests/higgs_tts/higgs_tts_warm_bench_cases.json. You can use them as reference.

Current numbers:

default
Python: 3215.28 ms, C++: 2584.43 ms

multi_session with 3 requests
Req 0: Python 3155.95 ms, C++ 2528.38 ms
Req 1: Python 1876.53 ms, C++ 1913.05 ms
Req 2: Python 1763.17 ms, C++ 1761.59 ms
Total: Python 6795.65 ms, C++ 6203.03 ms