Support for using step level advantages from verifiers

[eligotts]

This PR adds support for reading verifiers defined step level advantages. In order for prime-rl to read these, we need:

• flag to be set in config
• verifiers to signal this through state, pr here: send step level advantages to prime rl verifiers#941
• for each trajectory step, the completion_advantages needs to be the same length as completion_ids

if any of those aren't set/fail, we fallback to distributing the prime-rl calculated advantage to each token

---

Note

Medium Risk
Touches the RL training data path (transport schema + advantage packing), so mismatched lengths or missing fields could break training; safeguards exist via validation/asserts and fallback behavior.

Overview
Adds optional support for verifier-provided step/per-token advantages via new orchestrator.use_verifier_step_advantages config, plus plumbing to request a use_verifiers_advantages state column in rollout generation.

When enabled and verifier data is present/valid, interleave_rollout populates TrainingSample.completion_advantages from per-step completion_advantages (padding prompt-in-completion segments with zeros); otherwise the orchestrator backfills completion_advantages by repeating the computed rollout advantage per token. This also replaces the transport field advantage with completion_advantages, updates batch packing to consume per-token advantages (prompt tokens forced to 0), and adjusts/extends unit tests accordingly.

^{Written by Cursor Bugbot for commit 2e94744. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with Cloud Agents, enable Autofix in the Cursor dashboard.}