Docs + Environment pattern: RLHF

[bxyu-nvidia]

Use cases, pain points, and background

Description:

Design: We probably need to make some generic reward model client that can be shared infra for all RLHF environments.

Out of scope:

Acceptance Criteria:

• [ ] Gym spins up a reward model locally like in the local vLLM model flow
• [ ] Replicate the current Nemotron RLHF process