Add EngGPT MoE model support

[robertobissanti]

Summary

This PR adds initial support for the enggpt_moe architecture used by engineering-group/EngGPT2-16B-A3B.

The model is a decoder-only MoE language model with:

• 24 decoder layers
• hidden size 2880
• 32 attention heads
• 4 key/value heads
• explicit head_dim = 128
• Q/K RMSNorm inside attention
• RoPE with rope_theta = 1000000.0
• 64 experts per MoE layer
• top-8 expert routing
• SwiGLU experts
• untied lm_head

The implementation is based on the existing Mixtral/SwitchGLU infrastructure, adapted to match the EngGPT MoE checkpoint structure and routing logic.

Motivation

The EngGPT2-16B-A3B model cannot currently be loaded or converted with mlx-lm because its config.json declares:

"model_type": "enggpt_moe"