Performance: Sparse adam

[ArthurBrussee]

Gradients are already calculated as sparse gradients. As a very final step, these are then made dense again and fed to the optimizer.

We'd ideally have a sparse adam optimizer. It'd be great if this was a re-usable library for Burn, but more likely it'd mean just implementing this as a custom kernel.

[ArthurBrussee]

After some discussions, I think #38 is the more fruitful direction. Sparse adam is not a big win as momenta will be dense anyway. Taking multiple images/step will also make the gradients more "dense"