feat(rollout): collect complete greedy episodes for ReMax baselines

[pty12345]

Summary

• Collect a complete greedy agent episode from an independent environment reset alongside sampled episodes.
• Pair sampled steps with the greedy episode's undiscounted reward by source UID.
• Preserve episode termination and truncation metadata, with tests covering rollout collection, environment adapters, reward pairing, and tensor layout.

Scope

This PR provides the rollout-collection foundation for ReMax. It does not yet implement algorithm.adv_estimator=remax, advantage computation, or the ReMax actor update.

Validation

• 28 tests passed in the full CPU environment.
• The local test suite passed, with dependency-gated tests skipped where the full training stack is unavailable.