bugfix: fix mbox model(qwen2.5) multi round core in xattention

[DragonFive]

Summary

This PR fixes the REC multi-round core path in xattention for the mbox (Qwen2.5-related) model flow by migrating the missing KV-cache
attachment logic.

Problem

In REC multi-round mode, the model forward path did not fully attach multi-round cache tensors (full_k/v and unshared_k/v) into
attention metadata for affected model implementations.
This could break or degrade multi-round behavior in xattention.

Changes

• Extend REC model-type detection to include qwen3_moe.
• In LlmModelImplBase forward:
  • Read LlmRecMultiRoundParams only when REC multi-round mode is enabled and params are present.
  • Add per-layer size checks for full_k_caches, full_v_caches, unshared_k_caches, and unshared_v_caches.
  • Attach the corresponding cache tensors into attn_metadata before each layer forward.
• Apply the same REC multi-round cache attachment logic in Qwen3MoeModelImpl forward.

Impact

• Enables correct multi-round KV-cache wiring for xattention in the affected REC path.
• No behavior change for non-REC-multi-round execution paths.

Files Changed

• xllm/core/common/rec_model_utils.h
• xllm/models/llm/llm_model_base.h
• xllm/models/llm/qwen3_moe.h

Notes

• This PR focuses on wiring/attachment correctness for multi-round REC cache metadata.
• Follow-up cleanups (for example, redundant null checks) can be done separately to keep this bugfix focused.

[gemini-code-assist]

Code Review

The pull request introduces support for qwen3_moe models in the is_llmrec_model_type function and integrates multi-round caching parameters (LlmRecMultiRoundParams) into the LlmModelImplBase and Qwen3MoeModelImpl forward passes. This change is crucial for enabling multi-round core functionality in xattention for these models. The added includes and logic for handling llmrec_params are appropriate for the stated bugfix.