feat(gym): persist stable rollout attempt identity

[ananthsub]

Summary

• assign one stable logical _ng_rollout_id to each prompt sibling
• allocate numeric _ng_attempt_index values only when a sibling is dispatched
• preserve attempt allocation across partial row redispatch, outer retries, and rollout recovery
• derive Gym capture keys as <logical> for attempt 0 and <logical>-a<N> for later attempts
• bump rollout recovery state to schema v3 and reject unmigratable UUID-based v2 state with guidance

This is the first PR in the NeMo Gym checkpoint integration stack. It intentionally excludes the NemoGym checkpoint participant and SingleController orchestration.

Related Gym tracking issue: NVIDIA-NeMo/Gym#3024

Supersedes fork-based draft #4085.

Validation

• combined rebased stack tests on current upstream/main: 245 passed
• focused rollout identity, manager, recovery, failure, redispatch, and NemoGym tests before restacking: 202 passed
• native redispatch suite: 25 passed
• Ruff and formatting: passed
• Pyrefly: 0 errors in the prior focused run

Stack

Next: #4093 — add the NemoGym actor checkpoint participant.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.