[Feature] Add support for in-flight weight updates

[lewtun]

In-flight weight updates were proposed by PipelineRL and allow one to stay near on-policy without tanking throughput:

It was also validated in the ScaleRL paper, that adopting in-flight weight updates dramatically speeds up training by factors of 2-3x over conventional off-by-k training:

I poked around the torchforge codebase and did not see this implemented yet, but think it would be a high impact feature for the open source community to include :)