feat: add onerec 3b performance optimization and support old model.

[DragonFive]

Summary

• align OneRec 3B NPU runtime with xllm_rec
• keep legacy and 3B scaling paths gated in the NPU runtime
• update xllm_atb_layers submodule to merged master commit 96d3deb2

Notes

• branch keeps two commits intentionally: one for 3B runtime alignment, one for legacy/3B gating and submodule linkage
• xllm_atb_layers changes are already merged into git_code master

[gemini-code-assist]

Code Review

This pull request enhances the NPU OneRec block layer by introducing ACLNN-based attention linear support, a decoder prefill-only execution mode, and improved handling for fused FFN and MoE expert weights. It also adds a compatibility path for specific checkpoint prefixes and updates the MoE attention mask logic. Review feedback highlights a critical memory safety issue where a local tensor is referenced via a dangling pointer in the variant pack. Furthermore, multiple style guide violations were noted, specifically regarding the use of auto for simple types, missing vector reserve() calls, and the need for parameter annotations on constant arguments.

[DragonFive]

/gemini review

[gemini-code-assist]

Code Review

This pull request updates the OneRec NPU implementation to support ACLNN-based attention linear operations, introduces a prefill-only execution path for decoders, and enhances weight loading and merging for MoE layers, including support for fused expert weights and dynamic quantization. Feedback highlights a critical thread-safety concern regarding the storage of request-specific data as member variables in the layer implementation, which could lead to data corruption in concurrent environments. Additionally, several style guide violations were noted, specifically the improper use of 'auto' for simple types and the use of plain 'int' instead of fixed-width integers.