fix(vibevoice): preserve quantization metadata in sanitize() for quantized model loading

[korale77]

Context

Loading quantized VibeVoice TTS models (8/6/5/4-bit from mlx-community) fails with:

ValueError: Expected shape (151936, 896) but received shape (151936, 224)
for parameter language_model.embed_tokens.weight

The fp16 variant loads fine. This is the same class of bug fixed for fish_qwen3_omni in #584.

Description

sanitize() filters out weight keys not found in the model's current parameter shapes. At sanitize time the model hasn't been quantized yet, so quantization metadata keys (.scales, .biases) don't exist in the model and are silently dropped.

Later, apply_quantization() checks f"{p}.scales" in weights to decide which layers to quantize. Since sanitize removed the scales, it skips quantization entirely. The model stays full-size but the weight file has quantized shapes, causing the mismatch.

The fix preserves .scales and .biases keys through sanitization, matching the existing pattern in chatterbox/s3gen.

Changes in the codebase

• mlx_audio/tts/models/vibevoice/vibevoice.py: In sanitize(), preserve quantization metadata keys (.scales, .biases) that aren't in curr_shapes, instead of dropping them. This allows apply_quantization() to find the metadata and properly quantize the model before weight loading.
• mlx_audio/tts/tests/test_models.py: Add test_sanitize_preserves_quantization_metadata to verify .scales and .biases keys survive sanitization.

Changes outside the codebase

None.

Additional information

• Same root cause as fix(fish_qwen3_omni): support loading quantized/converted models in sanitize() #584 (fish_qwen3_omni sanitize fix), different fix approach because vibevoice uses shape-based filtering while fish_speech uses prefix-based filtering.
• The .endswith((".scales", ".biases")) pattern matches the existing approach in chatterbox/s3gen/s3gen.py.
• Tested with mlx-community/VibeVoice-Realtime-0.5B-8bit.

Checklist

• Tests added/updated
• Documentation updated
• Issue referenced (e.g., "Closes #...")

[lucasnewman]

Thanks!