Align aggregate metrics with persisted rollouts

[fallintoplace]

Summary

• aggregate rollout metrics from the rows written to the main rollouts JSONL only
• keep sidecar failures and _ng_no_persist attempts out of runtime /aggregate_metrics
• preserve already-persisted rows when resuming from cache
• add regression coverage for successful, sidecar-failure, no-persist, and resumed rollouts

Root cause

run_from_config() appended every result to the in-memory results list before the persistence gate, then passed that full list into _call_aggregate_metrics(). That let runtime aggregation include rows that never landed in output.jsonl, so gym eval run could disagree with gym eval aggregate.

Validation

• uv run --extra dev python -m pytest tests/unit_tests/test_rollout_collection.py

[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.

[marta-sd]

Thanks you for your contribution @fallintoplace !

I added one small comment. Also please sign your commits as described in contributing guideliness so we can accept your PR :)

[marta-sd]

/ok to test f55751e

[fallintoplace]

Somehow the CI fails. I can have a look later today.