feat: add flags for multi_round pipeline to return logprob.

[DragonFive]

PR Description

Summary

This PR adds optional REC logprob output for pure-device multi-round beam results on top of main.

Background

main already has the beam ranking order fix (descending by beam logprob).
The remaining gap was REC multi-round output logprobs in generate_multi_round_output.

What changed

1. Added a new global flag:

• FLAGS_output_rec_logprobs (default: false)

2. Updated REC multi-round output behavior in SequencesGroup::generate_multi_round_output:

• If FLAGS_output_rec_logprobs == true:
  • Return a full logprobs list aligned with token_ids
  • Fill missing per-token logprobs with last_lps[beam_idx]
• If FLAGS_output_rec_logprobs == false:
  • Do not return out.logprobs

3. Kept beam ranking logic unchanged.

Files changed

• xllm/core/common/global_flags.h
• xllm/core/common/global_flags.cpp
• xllm/core/framework/request/sequences_group.cpp

Behavior matrix

• output_rec_logprobs=false (default): no REC multi-round out.logprobs
• output_rec_logprobs=true: token-aligned out.logprobs, filled by final beam logprob

Compatibility / performance

• Default is false, so existing behavior remains unchanged unless explicitly enabled.
• Enabling the flag increases output payload and per-output construction work proportionally to token count.

Validation

• Verified symbol wiring and code paths (DECLARE/DEFINE/usage).
• Verified only target files were modified for this feature.
• Full build/test was not run in this patch.

[gemini-code-assist]

Code Review

This pull request introduces a new flag, output_rec_logprobs, to enable the output of token-aligned log probabilities for multi-round recommendation pipelines. While the feature is a good addition, the current implementation has a couple of issues. First, it uses an incorrect default value for log probabilities in an edge case, which could lead to misleading results. Second, there is a performance concern due to inefficient token decoding within a loop. I have provided specific comments and code suggestions to address these points.