Add "Training" Section to Documentation

[cwing-nvidia]

Background

Users have asked if they could "use Gym with external models for DPO data collection" as if this was a special case, when it's actually a core use case that NeMo Gym supports.

Problem

We have a SFT/DPO tutorial but the information needs to be featured more prominently in our docs.

Acceptance Criteria

• [ ] Create a new "Training" section in documentation
• [ ] Explicitly show how NeMo Gym supports:
  • RL training with NeMo RL
  • SFT data collection
  • Preference data collection for DPO
• [ ] Clarify that Gym can be used with any OpenAI-compatible endpoint

Priority

High - common user question, validates work already in progress