feat: add Harbor integration

[grace-lam]

Adds the full Harbor agent integration in NeMo Gym for multi-turn RL training workflows, including custom agent/env wiring, rollout conversion, and operational docs.

Scope

This PR covers the full responses_api_agents/harbor_agent/ package:

• app.py (Harbor job orchestration + response shaping)
• utils.py (trajectory → NeMo Gym output conversion)
• custom_agents/ (Terminus-2 NeMo Gym compatibility path)
• custom_envs/singularity/ (HPC-oriented execution environment)
• configs/harbor_agent.yaml (default agent config)
• tests/test_app.py (integration/unit behavior checks)
• README.md