feat: add conditional diffusion model for 3D molecule structure generation

[georgosgeorgos]

• Draft that introduces GeoDiff (https://arxiv.org/abs/2203.02923) in src/gt4sd/algorithms/generation/diffusion/geodiff. GeoDiff is an equivariant diffusion model for 3D molecular structure (conformations) generation.
• The "prompt" for this model is a 2D representation.

[georgosgeorgos]

It looks great to me!

The only thing I would still consider is keeping the support for diffusers<0.6.0. I did this in #157. If that would be merged first, the only thing that would need to be adapted here is the return type of geodiff, either item['sample'] or item.images, dependent on the version. This would also get rid of the if/else in diffusion/core.py which would be replaced by a version-specific if/else (see #157 )

We still need something like if/else for diffusion/implementation.py. In geodiff we return a rdkit.Mol object and using item.images can be very confusing.

[georgosgeorgos]

In the get_targets_description I list the keys of the dict. Do we want a description for each key?

[drugilsberg]

In the get_targets_description I list the keys of the dict. Do we want a description for each key?

No it's enough to report the schema of the dictionary and report an high level description. It is useful to understand only the schema of the target.