Upload MLX model via mlx-forge

Update split_model.json for multi-variant support

fix: transpose vocoder buffers (filter, basis) and upscaler kernel to MLX channels-last layout

Rename upscaler files with version (x2_v1_1, x1_5_v1_0, x2_v1_0)

Add spatial upscalers (x2, x1.5) and temporal upscaler (x2) for multi-stage pipelines

fix: pre-convert ConvTranspose1d (ups) weights to MLX layout (O,K,I)

Upload MLX model via mlx-forge