[train][Fix] Fix process group creation with vLLM for run_engines_locally=false

[SumanthRH]

What does this PR do?

Fixes process group creation with vLLM for run_engines_locally=false. Currently, weight syncing with run_engines_locally is broken, and the remote engine example https://github.com/NovaSky-AI/SkyRL/blob/main/skyrl-train/examples/remote_inference_engine/run_remote.sh fails with NCCL error : invalid argument

This is because of our custom process group creation logic - I suspect this broke after a PyTorch upgrade. It is recommended to use vLLM's StatelessProcessGroup abstraction for creating multiple process groups without polluting global state.

This PR only modifies the vLLM + NCCL backend codepath to use the new process group creation logic. Fixes for Sglang will follow in a follow-up PR

[gemini-code-assist]

Code Review

This pull request introduces a fix for creating process groups with vLLM when run_engines_locally=false. A new utility stateless_init_process_group is added to leverage vLLM's StatelessProcessGroup, avoiding conflicts with the default torch distributed process group. This is conditionally used when the weight_sync_backend is nccl and the generator backend is vllm. The changes are consistently applied across the vLLM inference engine and the DeepSpeed/FSDP workers. The overall logic is sound, but I found a critical typo in the FSDP worker that needs to be addressed.