Support broadcasting just lora weights and drop support for merged weights

[Jackmin801]

Note

Adds adapter-only filesystem weight broadcasting and runtime LoRA loading, wires LoRA names through orchestrator/scheduler, removes merged-weight path, and updates tests/CI accordingly.

• LoRA broadcasting and loading:
  • Add adapter_only option to weight broadcast configs (trainer/rl/config.py); filesystem broadcaster now saves only adapters and writes adapter_config.json; NCCL path disallows adapter-only.
  • Introduce get_adapter_state_dict and adapter filename handling in trainer/weights.py; drop merged-LoRA gathering path; simplify gather_weights_on_master.
  • Pass lora_config into broadcaster setup and persist with adapter saves.
• Orchestrator & Scheduler:
  • Add OrchestratorConfig.lora_name and thread through Scheduler; use LoRA name as model_name after updates.
  • update_weights now accepts lora_name and uses new /v1/load_lora_adapter; avoid base reload when using LoRA.
• Inference (vLLM server):
  • Inject hack to bypass adapter already-loaded checks to allow in-place LoRA updates.
• Trainer:
  • Broadcast weights each step using adapter_only flag; checkpoint manager unchanged but aligned with new adapter flow.
• Tests/CI:
  • Add integration tests for LoRA RL with dynamic adapter loading; separate LoRA tests in CI; capture vLLM/stdout logs for debugging.

^{Written by Cursor Bugbot for commit 7857428. This will update automatically on new commits. Configure here.}

[mikasenghaas]

very clean, i love it. nice that we decouple the full vs adapter only weight saving as well, has been bugging me for a while how entangled these were