Separate single/multinode and enable CTX

[amd-wsung102]

Motivation

This PR continues Amir's commit cde5f3d and enables CTX for multinode cases using the new template variables.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[amd-wsung102]

The changes made in this PR include:

• Using the existing code changes in this commit.
• Enabling CTX in internode_ll.cu for multinode use cases. A multinode case is defined when the newly added template variable multinode is true, which depends on the number of ranks during runtime dispatch.

[RichardChamberlain1]

Needs some general tidy up.

I ran the a single node test and it hung on inter-node, see here...

https://ml-ci-internal.amd.com/job/DeepEP/job/Experimental/101/console.

Has this been stress tested? And if so what docker image did you base it off?

[amd-wsung102]

The docker image is based on rocm6.3.4_ubuntu24.04_py3.12_pytorch_release_2.4.0. This has not been stress tested. I am currently trying to stress test it on OCI but OCI is showing this error message permission denied while trying to connect to the docker API at unix:///var/run/docker.sock.

I will continue to try running it on OCI.