fix(convert): validate MLX quantization modes

**Because** the shared converter filtered every quantization candidate by 64-element groups even when MXFP and NVFP require 32- and 16-element groups, so valid layers could be silently left dense and invalid mode combinations reached MLX only after model loading.

**This commit** centralizes the public MLX mode contracts, validates group-size and bit-width combinations before mutation, applies the selected group to layer eligibility, documents the exact matrix, and exercises affine, MXFP4, MXFP8, and NVFP4 conversion/load/generation paths for MiniMax Music 3.\n\nCo-Authored-By: Codex <noreply@openai.com>