feat: add onerec model implement[2/N]

[DragonFive]

Summary

This PR is the second step of the OneRec rebaseline stack, following #970
It focuses on NPU block-layer wiring and OneRec runtime input/plumbing hardening on top of current main, while keeping existing non-
OneRec model paths stable.

What's included

• Add OneRec block-layer contract:
  • xllm/core/layers/onerec_block_layer.h
• Add NPU OneRec block-layer implementation:
  • xllm/core/layers/npu/npu_onerec_block_layer_impl.h
  • xllm/core/layers/npu/npu_onerec_block_layer_impl.cpp
• Wire new OneRec layer into build targets:
  • xllm/core/layers/CMakeLists.txt
  • xllm/core/layers/npu/CMakeLists.txt
• Extend OneRec request input handling in rec_master:
  • Validate supported tensor names (sparse_embedding, decoder_context_embedding)
  • Validate dtype/shape/content size
  • Validate required tensors and hidden-size consistency
  • Convert accepted tensors into MMData
• Adapt runtime wiring for OneRec worker path:
  • RecWorkerImpl OneRec model init and forward path integration
  • Pipeline-side cleanup/fixes to align with current worker/executor usage
• Scheduler guard fix:
  • Restrict REC multi-round scheduling condition to RecType::kLlmRec
• Follow-up fixes:
  • onerec_args vocab/tokenizer argument fix
  • Device placement fix in NPU OneRec block layer (device_-based move path)
  • Build-error and merge-conflict cleanup in touched runtime code

Scope / Restrictions

• This PR is primarily wiring and integration for OneRec on NPU path.
• It does not introduce new non-OneRec runtime behavior.
• The OneRec NPU block layer remains integration-level logic (no large algorithmic kernel expansion in this PR).

Backward Compatibility

• Existing llm / vlm / embedding / non-OneRec REC flows are preserved.
• OneRec-specific code paths are activated only for OneRec model/type selection.

Testing

• Build-level validation for touched targets (layers, runtime, scheduler, models/rec).
• Source-level checks for OneRec input validation branches in rec_master.
• Smoke checks for OneRec worker forward path (encoder/decoder branch handling and device placement).
• Regression check that non-OneRec pipelines remain unchanged.

[gemini-code-assist]

Code Review

This pull request introduces the wiring for a new onerec_block_layer for NPU, including the abstract base class, NPU implementation, and build system changes. The implementation of NpuOneRecBlockLayerImpl::forward appears to be a placeholder, which is expected for a wiring PR. I've found a critical issue in the tensor device placement logic within the forward method that could lead to runtime errors. Please see the specific comment for details.