feat: support DeepSeek-V4 in the Python model executor

[0moyi0-2024]

Summary

• bridge DeepSeek-V4 cache and attention metadata into the Python executor
• expose the NPU DSA, QLI, compressor, partial RoPE, hyper-connection, and grouped MoE operators to Python
• add the DSA metadata builder and eager attention backend aligned with the native implementation
• add the DeepSeek-V4 Python model, weight loading, MoE path, and registry integration
• reuse native tensor-parallel and expert-parallel collectives instead of creating a second Python HCCL communicator

This PR intentionally covers the 8-card eager path only. ACL graph, graph double buffering, CP, PD disaggregation, and host offload are excluded from this change.

Validation

• Python regression suite: 111 passed
• Python syntax and git diff --check: passed
• modified C++ translation units: compiled
• full NPU build and link, including all_tests: passed
• torch_npu_ops Triton build tests: 212 passed
• 8-card HCCL startup with /mnt/sfs_turbo/models/DeepSeek-V4-Flash-w8a8-mtp: passed
• real Python eager weight loading, KV-cache initialization, prefill, and decode request: passed
• deterministic Python eager vs native eager comparison: generated token ids and per-token top-5 logprobs matched exactly for the validation request

Runtime configuration

• model_impl=python
• enable_graph=false
• python_graph_backend=off
• npu_kernel_backend=TORCH
• ep_size=8, dp_size=1
• prefix cache disabled for the baseline validation