Prepare self-contained speech bundles for 0.5.2

[benjipeng]

What changed

• package Step Audio EditX as one self-contained runtime bundle
• make Fish S2 load its bundled codec and tokenizer assets only
• default Granite Speech scripts to the published int8 checkpoint
• fix Cohere conversion so eligible MLX layers are actually quantized
• bump the package version from 0.5.1 to 0.5.2

Why

Several local conversion layouts differed from their published runtime layouts. Cohere also passed the upstream BF16 state dictionary into the conversion-time quantization selector, so no modules were quantized even though the artifact metadata said int8. These changes make the converters, loaders, local bundles, and published Hugging Face artifacts agree.

Impact

Runtime loaders now consume only the assets inside the supplied model bundle. Existing public APIs remain unchanged. The corrected Cohere artifact contains 517 affine int8 Linear/Embedding modules while convolution and normalization paths retain floating-point precision.

Validation

• pytest tests/unit/: 920 passed
• Step Audio real voice-clone inference passed from the self-contained bundle
• Fish S2 generated a 22-second waveform from the self-contained bundle and passed ASR verification
• Cohere int8 produced a transcript identical to the BF16 reference on the 22-second test clip
• Dots int8 strict checkpoint loading passed without the deleted local base artifacts
• Ruff checks passed for every modified Python file
• CPU-only build produced mlx_speech-0.5.2-py3-none-any.whl and mlx_speech-0.5.2.tar.gz