mlx-gemma4: SwitchGLU experts + sanitize fixes

mlx-gemma4: add GGUF loading + Q8→4-bit conversion

mlx-gemma4: KV sharing is ACTIVATION reuse, not weight sharing

mlx-gemma4: add KV sharing + PLE, E2B loading progress

Add mlx-gemma4 standalone package

Gemma 4 Q8 benchmark: 0.27 tok/s on 16 GB (llama.cpp: 0 tok/s)

GEMMA 4 WORKS! RMSNorm fix: raw weights, no +1 offset

Add Gemma 4-26B MoE to tiny-bit-terminal compare + models

Add Gemma 4-26B MoE to tiny-bit-terminal model options

Gemma 4: architecture confirmed correct, need expert FFN fix

Gemma 4: add post_feedforward_layernorm after dense+moe combine

Gemma 4: revert to (1+w) norms, debug layer-by-layer

Gemma 4: fix RMSNorm for GGUF weights + attention fixes

Fix missing post_attention_layernorm in Gemma 4 model

Gemma 4 MLX engine: forward pass works, needs weight mapping fix