Update rocshmem and enable GDA

[erieaton-amd]

This enables the RDMA/GDA support that was added to rocshmem. It's working on a single node with an MLX5 card. This WIP has not yet been tested on separate machines.

[CLAassistant]

All committers have signed the CLA.

[wenlei-bao]

Thanks for the contribution. @erieaton-amd @drprajap Is this PR ready?
BTW we just update Triton-distributed/Triton, so you may want to rebase.

[KnowingNothing]

We have updated the code. If you have any problems in refactoring your PR, feel free to leave comments : )

[erieaton-amd]

I refactored the PR. However, I am having some trouble getting even the main branch to work now.

[erieaton-amd]

I have the tests test_put_signal.py and 03a-inter-node-allgather.py reporting that they passed, but they crash immediately afterward. There appears to be some memory corruption issue.

[wenlei-bao]

@erieaton-amd @drprajap Seems CI failed, can you please take a look?

[erieaton-amd]

The code right now only works if the machine has a Mellanox card set up. This is because the rocshmem backend is hard coded into the bitcode, which has to be cleaned up somehow so it can use the IPC backend also. Right now I'm trying to figure out why the tests I wrote are now failing on the machine that does have a Mellanox card.

[erieaton-amd]

Ok, the tests seems to be working again. I've enabled a dispatch feature in rocshmem that should select the right backend in the bitcode, so maybe the CI will work now.

[erieaton-amd]

@wenlei-bao I am investigating an issue that the function sleep_async doesn't work in newer versions of ROCm. This doesn't appear to block this work though, so the patch can be run through the CI and reviewed.

[erieaton-amd]

The bitcode was missing a file, try the CI again.