feat: enable torch_npu graph mode for Qwen-3 dense with TP support.

[yingxudeng]

No description provided.

[yingxudeng]

Summary of Pending Items for this PR

This PR is a work in progress. The following items need to be completed before it's ready for final review:

• Compile the new libtorch_npu.so and update the related Docker images.
• Refactor the WordEmbedding class:
  • Rename the legacy WordEmbedding implementation to free up the class name.
  • Ensure the new WordEmbedding implementation can share unified logic with MLU and GPU backends.
• Apply the same refactoring (as point 2) to the LMHead component.
• Investigate and fix the widespread compilation failures in the test files.
• Validate that the changes do not break existing model inference pipelines.
• Review and update the default values for flags like ENABLE_NATIVE_NPU and USE_NPU_TORCH (modify or remove as needed).
• Analyze and optimize #if preprocessor directives for potential consolidation.
• Address other minor pending items and technical debt.

[XuZhang99]

we can get num_heads and num_kv_heads form th shape of query and key.

[yingxudeng]

In the file xllm_2/xllm/core/layers/common/attention.cpp, you’ve pre-shaped the query and key parameters before passing them. However, different platforms may require slightly different shapes after the view operation. It might be more flexible to pass the original query, key along with num_heads_ and num_kv_heads_ parameters, and then perform the view operation inside the batch_prefill method. This approach would provide better cross-platform compatibility and clearer parameter handling.

[yingxudeng]

Could you please help review this PR when you have a moment? 🙏 @yq33victor @XuZhang99