feat: Add dedicated lora mode to Megatron backend

[vivekkalyan]

Summary

This adds dedicated lora mode to the Megatron backend so inference and training can run on separate GPUs in parallel.

In dedicated mode, ART now keeps a dedicated vLLM server on the inference GPU and updates LoRA adapters in place after training steps. That makes the Megatron flow match the dedicated-serving model we already use elsewhere instead of treating training as a blocking operation.

What this enables

• Use Megatron with separate trainer and inference GPUs in dedicated mode
• Keep rollout serving on the inference GPU while Megatron trains on the trainer GPU
• Advance served model steps by reloading the next adapter into the running vLLM server

Validation

• Unit coverage:
  • tests/unit/test_megatron_dedicated.py
• 2-GPU smoke on a fresh H200 machine:
  • trainer on GPU 0
  • inference on GPU 1
  • base model Qwen/Qwen3-30B-A3B-Instruct-2507
  • dedicated lora mode completed two real train steps and advanced the served model from @0 to @2