Optimize ACE-Step: NLC VAE, compiled decode, LoRA support

[fspecii]

Summary

• VAE rewrite to NLC format: Replace WeightNormConv1d/WeightNormConvTranspose1d with nn.Conv1d and FastConvTranspose1d. Weight-norm parameters (weight_g, weight_v) are fused into regular weights at load time via sanitize(), eliminating per-forward-pass normalization overhead.
• Compiled VAE decode: mx.compile(model.vae.decode) with auto-conversion to mlx_weights.safetensors on first load (skips PT→MLX conversion on subsequent runs).
• mx.fast.scaled_dot_product_attention: Replaces manual attention (matmul → mask → softmax → matmul) with MLX's fused kernel.
• Simplified turbo diffusion: Single-pass inference without CFG/APG (turbo model was distilled without guidance). Removes ~80 lines of unused guidance code.
• LoRA adapter support: load_lora() / unload_lora() with weight fusion (W + scale * (alpha/r) * B @ A), base weight backup/restore for hot-swapping adapters.
• Quantized 5Hz LM variants: Added 0.6B-8bit and 0.6B-4bit model IDs for lower-memory language model inference.
• Music metadata: bpm, keyscale, timesignature parameters forwarded to prompt formatting.
• Model loading: custom_loading class attribute + acestep remapping for clean integration with mlx_audio.utils.base_load_model.

Test plan

• Generate 30s instrumental with --model ACE-Step/ACE-Step1.5 — verified output WAV
• Generate 30s track with lyrics and vocal_language param
• Generate with bpm/keyscale/timesignature metadata params
• Verify mlx_audio.tts.load() API path works with custom_loading
• Generate 60s track — verified correct duration and stereo output
• Verify other TTS models (Kokoro, Spark) unaffected by utils changes
• Test with LoRA adapter loading/unloading
• Test quantized LM variants (0.6B-8bit, 0.6B-4bit)
• Test audio-to-audio tasks (cover, extract, complete)

[lucasnewman]

I don't see the weight norm conv changes -- did those get excluded? Also, it looks like you removed cfg entirely? Was it not needed even for existing models? If you could explain the intent/goal of the changes vs just the mechanical pieces, it would be helpful context.

[Blaizzy]

@lucasnewman

This PR will be closed in favour of #499.

I will add @fspecii as a contributor there.