feat: bridge DeepSeek V4 Python caches and metadata.

[0moyi0-2024]

Summary

• extend the Python layer-cache contract with the six DeepSeek-V4 DSA cache slots while preserving the existing five generic slots
• expose DeepSeek-V4 forward metadata through the standalone Python attention metadata view
• bridge all eleven KV-cache entries from the C++ Python executor and preserve existing expanded-decode and MLA index interfaces
• add compatibility tests for legacy five-slot and DeepSeek-V4 eleven-slot cache tuples

Test plan

• PYTHONPATH=$PWD pytest -q tests/python/test_model_executor.py -k 'NormalizeLayerCaches or BindKvCaches' (6 passed)
• PYTHONPATH=$PWD pytest -q tests/python/test_model_executor.py -k 'not data_parallel_rejects_non_cuda_graph_backend' (42 passed, 1 deselected; the excluded baseline test expects the pre-ACLGraph error text)
• NPU/Torch C++ syntax checks for py_attention_metadata.cpp and py_executor_impl.cpp with -Werror -fsyntax-only
• full python setup.py build --device npu is currently running in the local workspace

Scope

This PR provides the cache and metadata bridge required by the DeepSeek-V4 Python model. The model implementation, DSA backend, custom kernels, and graph integration will be submitted separately.

[yingxudeng]

python setup.py build test --device npu 推荐本地单测也要跑下，加test 参数