feat: add onerec in supported model docs and align rec utility style.

[DragonFive]

Summary

This PR fixes the Rec support documentation and aligns a few Rec-related code paths with the current project style. This PR will merge after #1051 is merged.

Changes

• remove Qwen2.5 and Qwen3 from the Rec section in:
  • docs/zh/supported_models.md
  • docs/en/supported_models.md
• keep OneRec as the only documented Rec model entry
• enable FLAGS_enable_rec_prefill_only = true; in xllm/c_api/internal/rec.cpp
• replace torch::Dtype with torch::ScalarType in xllm/core/util/utils.cpp
• replace std::runtime_error with LOG(FATAL) in the same Rec tensor conversion path to match existing code style

Why

• Qwen2.5 and Qwen3 do not support ILU in the Rec path, so the previous documentation was incorrect
• the rec.cpp flag should be enabled instead of left commented out
• utils.cpp should follow the existing repository convention for PyTorch scalar type naming and fatal error handling

Validation

• checked the Rec section in both Chinese and English support docs
• verified that xllm/c_api/internal/rec.cpp now enables FLAGS_enable_rec_prefill_only
• verified that xllm/core/util/utils.cpp no longer uses torch::Dtype or std::runtime_error in the updated path

Notes

• no build or runtime test was executed in this local macOS editing environment
• this PR only updates the targeted files and does not address other existing workspace changes

[gemini-code-assist]

Code Review

This pull request updates documentation for supported Rec models, enables a feature flag, and aligns some code with project conventions. The documentation and flag changes appear correct. However, replacing std::runtime_error with LOG(FATAL) in xllm/core/util/utils.cpp introduces a critical issue. While this aligns with some of the file's style, it breaks the error handling in xllm/core/distributed_runtime/rec_master.cpp, which uses a try-catch block to handle these errors gracefully. The new implementation will cause the process to terminate on invalid input, which is a significant regression for a server application. I've added comments with suggestions to revert these specific changes to use throw again.