fix(gemma4): drop KV-shared layer projections in sanitize

[Fox13]

Problem

Quantized Gemma 4 safetensors files ship k_proj/v_proj/k_norm weights for layers >= num_hidden_layers - num_kv_shared_layers, even though those layers reuse KV projections from an earlier layer and have no corresponding model attributes. Loading any quantized Gemma 4 model fails with:

ValueError: Received 126 parameters not in model:
model.layers.24.self_attn.k_norm.weight,
model.layers.24.self_attn.k_proj.biases,
...
model.layers.41.self_attn.v_proj.weight

This affects mlx-community/gemma-4-e4b-it-4bit (52k downloads) and other quantized variants.

Fix

Drop those tensors in Gemma4TextModel.sanitize(), the complement of #1158. That PR stopped mlx-lm from creating unused projection slots for KV-shared layers. This PR drops the incoming weights for those slots that still arrive from the checkpoint. Together they make quantized Gemma 4 models load cleanly.

m = _re.search(r"\.layers\.(\d+)\.self_attn\.(k_proj|v_proj|k_norm)\.", k)
if m and int(m.group(1)) >= first_kv_shared:
    continue

Testing

Tested on M4 MacBook Air 24 GB with mlx-community/gemma-4-e4b-it-4bit. Loads cleanly and runs at 18–25 tok/s across 500–4K token prompts. Before this fix it failed immediately with the ValueError above.