Benchmark: Add archive support (generation and training)

[NicolasHug]

This PR adds support for generating archive-based datasets (tar archives, pickles, torch.save) with different binary data storage (BytesIO, tensor). It also enables training on such archive-based datasets on the torchvision recipes.

I'm not sure if this is something in scope for merging or not. If anything, it's useful as a reference for our future discussions regarding these topics.

If this is in scope and you'd like to merge it, a few things can probably be simplified / cleaned up a bit. E.g. it also adds support for tinyimagenet and simplifies the logger output, but that's unrelated to the goal of this PR - I'm happy to clean that up if needed.

I'm currently running these benchmarks on the cluster, which we can discuss tomorrow in our meeting!