Guard MLA caches from KV quantization

[xxxkkw]

Summary

• Add an explicit MLACache marker for DeepSeek-style latent MLA caches.
• Route DeepSeek V3 and Kimi K2.5 through MLA caches, and mark the latent side of DeepSeek V3.2's CacheList as MLA-specific.
• Skip kv_bits conversion for caches that do not support generic KV quantization, while keeping ordinary KV cache quantization unchanged.

Reproduction

Before this change, a tiny DeepSeek V3 config with kv_bits=4, kv_group_size=32, prompt length 3, and prefill_step_size=2 fails on the second step with:

AttributeError: 'list' object has no attribute 'swapaxes'

The failure happens because generic KVCache.to_quantized() turns the MLA rope side channel into a quantized tuple, but the DeepSeek V3 MLA attention path still needs an array for k_pe.swapaxes(...).

After this change, the same repro generates successfully and the cache remains MLACache.

Environment

• macOS Darwin 25.4.0 arm64
• Apple M1 Max, 32 GB unified memory
• Python virtual environment with MLX / mlx-lm test dependencies

Test plan

• python -m unittest discover -s tests -p test_models.py -k mla_models_make_mla_caches
• python -m unittest discover -s tests -p test_models.py -k deepseek_v3_kv_bits_skips_mla_cache
• python -m unittest discover -s tests -p test_models.py -k deepseek_v3
• python -m unittest discover -s tests -p test_models.py -k quantized
• python -m unittest discover -s tests -p test_prompt_cache.py -k cache_with_generate
• Synthetic 1-layer DeepSeek V3.2 and Kimi K2.5 cache-forward check with MLACache

No throughput speedup is claimed here; this is a correctness guard for an invalid quantization path. Generic quantized KV cache coverage continues to pass.