feat: Add Vibevoice7B model support

[Talpik]

Context

This PR improves non-streaming VibeVoice-7B support in mlx-audio and makes it substantially more
practical on Apple Silicon.

The work had two goals:

1. align the MLX non-streaming inference path more closely with the original vibevoice-community/ VibeVoice implementation
2. improve deployment options by adding model-aware selective quantization and validating practical
   speed/quality tradeoffs

The implementation was tuned and validated against the original upstream repo on a MacBook Pro 16" (M4 Max, 48GB RAM).

Description

The non-streaming VibeVoice-7B path was adapted and corrected to better match upstream behavior,
including prompt handling, reference voice prefill, negative-branch logic, grouped-query attention
behavior, and acoustic encoder parity.

After parity was improved, the runtime was profiled and selective quantization support was added in a
model-aware way, instead of relying only on the generic “quantize everything quantizable” path.

The PR also makes selectively-quantized checkpoints self-describing by saving quantized component
metadata into config.json, and updates model loading so quantized packed weights can be reconstructed
correctly during reload.

Experimental preset layers, speculative CFG speed paths, and investigation-only CLI tuning flags were
intentionally left out of this PR to keep the public surface area small and stable.

Changes in the codebase

• improved non-streaming VibeVoice-7B inference parity with upstream
• fixed multi-speaker and control-flow related issues in the non-streaming path
• fixed grouped-query attention parity in the Qwen2 LM path
• fixed acoustic encoder causal/strided padding parity for reference prefill
• added VibeVoice-specific selective quantization policy
• added support for persisting quantized_components metadata during conversion
• fixed quantized checkpoint reload so packed quantized layers are reconstructed correctly
• cleaned the public mlx_audio.tts.generate CLI by removing VibeVoice-specific experimental tuning
  flags
• removed internal non-streaming preset wrappers and speculative CFG speed paths from the PR version
• expanded VibeVoice test coverage for parity-sensitive and quantized-loading behavior

Changes outside the codebase

No infrastructure, database, or external service changes were made.

The original vibevoice-community/VibeVoice repository was used as the behavioral reference for parity
checks and runtime comparisons.

Additional information

Main comparison setup:

• same reference speaker
• same text
• cfg_scale=1.5
• ddpm_steps=10

Main comparison matrix

Mode	Audio Duration	Processing Time	Speed Factor	Peak Memory	Notes
Original PyTorch full	25.07s	42.65s	0.59x	n/a	Quality reference
MLX full	26.00s	103.02s	0.25x	25.21GB	MLX quality baseline
MLX full Q6	25.60s	21.12s	1.21x	9.57GB	Best general speed/quality tradeoff
MLX selective LM + prediction_head Q8	25.73s	24.02s	1.07x	14.19GB	Strong accelerated-
full candidate

Model preparation

Before running the commands below, download the original upstream
vibevoice-community/VibeVoice weights locally.

Assume the original PyTorch checkpoint is available at:

• <ORIGINAL_VIBEVOICE_WEIGHTS>

1. Full MLX port

python -m mlx_audio.convert \
  --hf-path <ORIGINAL_VIBEVOICE_WEIGHTS> \
  --mlx-path <OUTPUT_FULL_MLX_PATH>

#### 2. Full-model quantized Q6

python -m mlx_audio.convert \
  --hf-path <OUTPUT_FULL_MLX_PATH> \
  --mlx-path <OUTPUT_Q6_MLX_PATH> \
  --quantize \
  --q-bits 6

#### 3. Selective accelerated-full mode (LM + prediction_head Q8)

MLX_AUDIO_VIBEVOICE_QUANTIZE_PREDICTION_HEAD=1 \
python -m mlx_audio.convert \
  --hf-path <OUTPUT_FULL_MLX_PATH> \
  --mlx-path <OUTPUT_LM_PRED_Q8_MLX_PATH> \
  --quantize \
  --q-bits 8

### Generation command

Use the same generation command for all three variants by changing only
the --model path.

python -m mlx_audio.tts.generate \
  --model <MODEL_PATH> \
  --ref_audio <REFERENCE_AUDIO> \
  --text "$(cat <TEXT_FILE>)" \
  --cfg_scale 1.5 \
  --ddpm_steps 10 \
  --output_path <OUTPUT_DIR> \
  --file_prefix <RUN_NAME> \
  --audio_format wav \
  --verbose

Examples for <MODEL_PATH>:

- full MLX: <OUTPUT_FULL_MLX_PATH>
- full Q6: <OUTPUT_Q6_MLX_PATH>
- selective LM + prediction_head Q8: <OUTPUT_LM_PRED_Q8_MLX_PATH>

## Checklist

- [x] Tests added/updated
- [x] Documentation updated
- [ ] Issue referenced (e.g., "Closes #...")

[lucasnewman]

@Talpik Thanks for the contribution!

Can we split this into two PRs to reduce the size of the changes needed to review?

1. The model updates, which should be verifiable against the reference implementation.
2. The selective quantization enhancements that you've added for the performance tuning.