Avoid DGL dependency

[jvdoorss]

The first installation of RFdiffusion turned out to be a bit complicated, the culprit being the DGL and other dependencies of the SE3Transformer package.

While most dependencies of SE3Transformer are in fact redundant to RFdiffusion (they are used only in the packages benchmarking scripts), a graph learning package like DGL is essential. However, DGL seems to work only up to torch==2.4.1 and appears to be no longer maintained.

I've created a minimal adaptation of SE3Transformer, bypassing DGL and slimming it down to the core components only here. Apart from needing a different format of Graph, the functionality is unchanged there.

This PR would allow to use this DGL-less SE3Transformer making the changes:

• convert DGL graph to the new Graph object
• adjust the imports to the slightly simplified architecture of the SE3Transformer package
• adjust installation instructions
• remove the included env folder

and non-essential changes

• updates to avoid warnings
• unsqueeze edge features only in the transformers caller (Str2Str) rather than in the graph creating function, this seems more intuitive.

As a result, the package is pip installable and works with the latest versions of python and torch. Not sure if this is of use to any other users or if there are no other solutions out there to obtain this already, but if so I figured it might be a useful addition to the repo?

Unfortunately, diffusion tests are failing both for main and this branch, in a similar way. Not much to be learned from that. 🤷‍♂️

[VKudlay]

Hey @jvdoorss. Happened to be looking into a DGL dependency yesterday and saw your note. If that package proved to be challenging, the approach adopted by PhysicsNeMo may be of interest (switching to pygeometric). Not my area of focus right now, but just dropping the ref in case it's useful.

[jvdoorss]

Thanks @VKudlay, great reference 👍

It contains many of the tricks I had to figure out in SE3Transformer. Looks like DGL is being phased out in PhysicsNemo as well... 🤔

[VKudlay]

I wonder why (though it looks like you already figured it out). Good luck!