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

- Rewrite VAE to native NLC format with nn.Conv1d and FastConvTranspose1d,
  fusing weight-norm (g*v/||v||) at load time instead of every forward pass
- Replace manual attention with mx.fast.scaled_dot_product_attention
- Simplify turbo diffusion to single-pass (no CFG/APG) matching upstream behavior
- Add compiled VAE decode (mx.compile) with auto-conversion to mlx_weights.safetensors
- Add LoRA adapter support (load/unload with weight fusion)
- Add quantized 5Hz LM variants (0.6B-8bit, 0.6B-4bit)
- Add music metadata params (bpm, keyscale, timesignature)
- Add acestep model remapping and custom_loading support in utils

Author: fspecii