MLX conversion of meta-models/Muse-Glimmer-30B at 6-bit, made on an M1 Ultra.

Not supported by any released mlx-vlm, so it won't load in LM Studio. Works in oMLX today. For mlx-vlm, install the open PR:

Built from #1839 (quantized embedding keeps embed_norm), on top of the port in #1838. Thanks to @Blaizzy and @johntdavies.