fix: handle 3D KV tensors in prefix cache

[Thump604]

Restack of #144 onto current main.

What changed:

• keep the existing 4D block-concat behavior intact
• add the missing 3D KV path for Qwen3.5-style (heads, seq, dim) caches
• store/use the correct sequence axis in block metadata
• fix reconstructed offset accounting for 3D states
• add a regression test that reconstructs a partial prefix from a 3D KV cache

Validation:

• PYTHONPATH=/Users/David/code/vllm-mlx-pr144-restack /opt/ai-runtime/venv-live/bin/python -m pytest /Users/David/code/vllm-mlx-pr144-restack/tests/test_paged_cache.py -q
• PYTHONPATH=/Users/David/code/vllm-mlx-pr144-restack /opt/ai-runtime/venv-live/bin/python -m py_compile /Users/David/code/vllm-mlx-pr144-restack/vllm_mlx/prefix_cache.py /Users/David/code/vllm-mlx-pr144-restack/tests/test_paged_cache.py
• PYTHONPATH=/Users/David/code/vllm-mlx-pr144-restack /opt/ai-runtime/venv-live/bin/python -m black --check --fast /Users/David/code/vllm-mlx-pr144-restack/vllm_mlx/prefix_cache.py /Users/David/code/vllm-mlx-pr144-restack/tests/test_paged_cache.py

Fixes #142. Supersedes #144.