Add MiniMax M3 model support (text-only)

[machiabeli]

Summary

• Adds minimax_m3_vl.py model module for MiniMax M3 (427B MoE, ~23B active params)
• Builds on the existing minimax.py (M2.7) with key additions for the M3 architecture
• Phase 1: text-only with dense attention — vision tower and MiniMax Sparse Attention (MSA) deferred to a follow-up

Architecture highlights

Feature	Implementation
Hybrid dense/MoE	First 3 layers dense MLP, rest MoE (dispatched via moe_layer_freq)
MoE routing	128 experts, top-4 sigmoid + e_score_correction_bias + shared expert
SwigluOAI	Custom gated SiLU with alpha=1.702 scaling + limit=7.0 clamping
GQA	64 heads / 4 KV heads with per-head QK norm (Gemma-style RMSNorm)
Partial RoPE	rotary_dim=64 (half of head_dim=128)
FP8 source	Block-128 dequantization in sanitize()
Tensor parallelism	Full shard() support for distributed inference

Weight mapping

Expert weights on disk are per-expert 2D tensors (w1/w2/w3), stacked into 3D SwitchGLU format in sanitize() — same pattern as minimax.py. The language_model.model.* HF prefix is remapped to match the nested LanguageModel > TextModel structure.

Test plan

• from mlx_lm.models.minimax_m3_vl import Model, ModelArgs imports cleanly
• ModelArgs.from_dict(config) parses the nested text_config correctly
• Forward pass on tiny config produces correct output shapes
• mlx_lm convert --hf-path MiniMaxAI/MiniMax-M3 -q --q-bits 4 completes
• Converted model generates coherent text
• Quantization predicates keep routing gate at 8-bit