Add support for main features of GFN-Diffusion

[josephdviviano]

https://github.com/GFNOrg/gfn-diffusion/tree/main

This paper introduces many new features in a library which we should aim to support:
https://arxiv.org/abs/2402.05098

The first main contribution of this paper is a unified library for diffusion-structured samplers. The library has a focus on off-policy methods (continuous GFlowNets) but also includes simulation-based variational objectives such as PIS. Using this codebase, we are able to benchmark methods from past work under comparable conditions and confirm claims about exploration strategies and desirable inductive biases, while calling into question other claims on robustness and sample efficiency. Our library also includes several new modeling and training techniques, and we provide preliminary evidence of their utility in possible future work.