Fix Gemma 4 KV-shared layers creating unused projections

[glyphVault]

Summary

• Gemma 4 E4B/E2B models share KV projections across later layers (num_kv_shared_layers). The Attention class was creating k_proj, v_proj, k_norm, and v_norm for all layers, but shared layers never use them — the forward pass routes KV from earlier layers via shared_kv.
• This caused load_weights(strict=True) to fail for any Gemma 4 model saved through transformers (save_pretrained), since transformers correctly omits these weights for shared layers. This affects all derivative models: fine-tunes, merges, abliterations, etc.
• Skip creating k_proj/v_proj/k_norm/v_norm for KV-shared layers, matching the transformers implementation.
• Add a defensive ValueError if a shared layer somehow receives no shared_kv at runtime.

Test plan

• All 8 existing gemma4 tests pass
• New test test_gemma4_kv_shared_layers_omit_kv_projections verifies shared layers don't create KV modules
• Verified forward pass produces identical top-5 logits vs transformers on OBLITERATUS/gemma-4-E4B-it-OBLITERATED
• Verified cached generation produces coherent output
• Formatted with black

🤖 Generated with Claude Code

[angeloskath]

Thanks great catch.