fix(sc): regenerate buffered prompts on resume

[mehraakash]

Summary

• honor checkpointing.load_replay_buffer=false in SingleController native TQ restores
• validate the native replay snapshot before removing canonical and staging rows
• reconstruct the exact buffered prompts from stable prompt_idx tags and queue full fresh generations on the restored policy
• keep this change separate from rollout telemetry and Slurm/Pyxis changes

Validation

• git diff --check
• Python compile check for changed implementation and test files
• Added a focused replay-free restore test. Local pytest collection is currently blocked because this checkout requires Python 3.13.14 and the available environment is Python 3.13.5 without the complete project dependencies; CI should run the full supported environment.

Context

This is the current-main forward port of the replay-free resume behavior used by the SWE-E2E convergence work. Current main uses native TQ snapshots plus rollout-recovery ledgers, so this intentionally replaces the historical pending_rollouts.pt implementation instead of mechanically rebasing it.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[mehraakash]

/ok to test 9f1bc3a