Add early-finalize continuation for truncated reasoning rollouts

[taivu1998]

Summary

This PR adds opt-in support for early-finalizing truncated long-form generations so we can reserve answer budget inside a single completion window.

Closes #267.

What changed

• add a workflow-level early-finalize helper that:
  • reserves a configurable tail budget from max_tokens
  • runs an initial generation with the reduced budget
  • if that generation stops due to length, optionally appends a synthetic suffix only when reasoning was cut off inside an unfinished <think>...</think> block
  • continues generation from token input for the reserved tail budget
• add token-in/token-out continuation support to the rollout stack and timing layer for the Verl path
• wire the new helper into the standard workflow implementations and the FinQA workflow
• add Step.response_mask so synthetic suffix tokens remain in completion_ids for continuity while being excluded from Verl loss
• update the Verl transform to preserve the explicit step-level response mask instead of always assuming an all-ones loss mask
• add config knobs for rllm.early_finalize
• add focused tests for early-finalize behavior and response-mask propagation

Design notes

• the feature is opt-in and defaults to disabled
• v1 stays tightly scoped to the workflow/Verl path that the issue targets
• synthetic answer forcing is narrow by design: for non-thinking truncated outputs we simply continue from the partial completion without injecting a prefix
• existing prompt-length guards remain unchanged; this only handles the case where a single completion runs out of response budget

Testing

• python -m pytest tests/engine/test_early_finalize.py tests/unified_trainer/test_verl_transform.py tests/rewards/test_math_reward.py -q
• ruff check rllm/engine/rollout/rollout_engine.py rllm/engine/rollout/verl_engine.py rllm/workflows/early_finalize.py rllm/workflows/timing_mixin.py rllm/workflows/single_turn_workflow.py rllm/workflows/multi_turn_workflow.py rllm/workflows/cumulative_workflow.py rllm/agents/agent.py rllm/experimental/verl/transform.py rllm/experimental/verl/__init__.py projects/finqa/train_finqa.py tests/engine/test_early_finalize.py tests/unified_trainer/test_verl_transform.py
• python -m py_compile rllm/engine/rollout/rollout_engine.py rllm/engine/rollout/verl_engine.py rllm/workflows/early_finalize.py rllm/workflows/timing_mixin.py rllm/workflows/single_turn_workflow.py rllm/workflows/multi_turn_workflow.py rllm/workflows/cumulative_workflow.py rllm/agents/agent.py rllm/experimental/verl/transform.py rllm/experimental/verl/__init__.py projects/finqa/train_finqa.py

[kylemontgomery1]

I lean towards this being implemented at the workflow level (e.g., SingleTurnWorkflowWithEarlyFinalize) instead of a global feature in rLLM. @listar2000 What do you think?

[listar2000]

Will take a look thx @kylemontgomery1 for pointing this to me.

[taivu1998]

Hi @kylemontgomery1, @listar2000, could you help review again? Thanks!