Description:

For reward model env #1026
The implementation from def score(): to the referenced section is largely overlapping with get_logprobs.

To reduce duplication and improve maintainability, extract the common logic into shared utilities or a helper function.

Keep behavior unchanged; this should be a pure refactor. Add/adjust unit tests to ensure parity.

Scope:

• nemo_rl/models/policy/dtensor_policy_worker_v2.py
• nemo_rl/models/policy/dtensor_policy_worker.py

Acceptance Criteria:

No functional changes; existing APIs and outputs remain identical.

score() and get_logprobs no longer duplicate core logic.