fix: MLA and sliding window KV cache accuracy

[stmatengss]

Summary

• Fix 15x overestimate for MLA models (DeepSeek V2/V3/V3.1, Kimi K2): Replace the fake n_kv_heads=16, head_dim=512 hack with proper MLA compressed KV parameters (mla: true, kv_lora_rank: 512, qk_rope_head_dim: 64). The kvBytesPerToken() function now uses the correct formula: (kv_lora_rank + qk_rope_head_dim + kv_lora_rank) * dtype_bytes per layer = 1088 values instead of 16384 (a ~15x correction).
• Add sliding window metadata to Gemma 3 models (27B, 12B, 4B): Add sliding_window: 1024, sliding_window_pattern: 6 params to expose hybrid attention properties in the model profile.
• Update DeepSeek V3 deployment preset to use corrected MLA parameters.

Technical details

MLA (Multi-head Latent Attention) compresses KV cache using low-rank projection:

• K cache per layer = (kv_lora_rank + qk_rope_head_dim) * dtype_bytes = (512 + 64) * 2 = 1152 bytes
• V cache per layer = kv_lora_rank * dtype_bytes = 512 * 2 = 1024 bytes
• Total per layer = 2176 bytes vs old formula's 2 * 16 * 512 * 2 = 32768 bytes

Test plan

• All 178 JS parity validation checks pass (node tests/validate_js_parity.mjs)
• Verify DeepSeek V3 671B KV cache estimate in the web calculator shows ~0.12 GB/1K tokens instead of ~1.9 GB/1K tokens
• Verify Gemma 3 models display sliding window metadata in model notes