feat: add dedicated inference data parallel support

[vivekkalyan]

Summary

Enable dedicated Unsloth inference data parallelism across multiple inference GPUs.

This PR keeps the dedicated architecture as a single API frontend process with runtime LoRA reloading, and adds DP sizing/guardrails for multi-GPU inference.

What changed

• Allow multi-GPU inference_gpu_ids in dedicated config validation.
• Add dedicated validation for:
  • engine_args.data_parallel_size == len(inference_gpu_ids) (if provided)
  • engine_args.data_parallel_size_local == len(inference_gpu_ids) (if provided)
• In dedicated vLLM subprocess startup:
  • default data_parallel_size and data_parallel_size_local to inference GPU count when >1
  • default distributed_executor_backend to "mp" when >1
  • reject api_server_count != 1
• Add/expand unit tests for dedicated config validation contracts.

Testing

ART-E with 2 inference GPUs