Fix Gemma 4 quantized per-layer projection loading

[spicyneuron]

While trying to run unsloth/gemma-4-E2B-it-UD-MLX-4bit, I hit a loading error:

ValueError: Unable to quantize model of type <class 'mlx_lm.models.gemma4_text.ScaledLinear'>

After cross-checking this against the Gemma 4 implementation in Transformers (constructor, projection path), I replaced the custom ScaledLinear wrapper used for per_layer_model_projection with a standard bias-free nn.Linear and moved the hidden_size**-0.5 scale into the projection path explicitly.

The math stays the same, but the layer now works with MLX's normal quantization and loading flow.

For reference, here's the same fix on mlx-lm: ml-explore/mlx-lm#1112

[Blaizzy]

LGTM, thanks!