adjoints for cumsum and cumprod

[kraftpunk97-zz]

Another attempt at #282 .

cumsum is just reverse(cumsum(reverse(∆, dims=dims), dims=dims), dims=dims), as suggested by @mcabbott .

With cumprod, I had help from @dhairyagandhi96 who asked me to refer to PyTorch's implementation

I've had no way to test my gradients except for comparing my results with that of torch.cumprod for same inputs, so I haven't added any tests yet

[kraftpunk97-zz]

Another thing that comes to mind... I've had trouble getting the adjoints to work if I define them directly for the functions. However, if I wrap them around another function (like my_cumprod or my_cumsum) and then define their adjoints and call the gradient on these wrapper functions, they work.

The error I get is a DimensionMismatchError. I have copied the stack trace I get for cumprod.

[DhairyaLGandhi]

Something seems to have messed up in this commit?

[kraftpunk97-zz]

It's probably nothing. When I open the file in Github's editor, the formatting seems fine.

[DhairyaLGandhi]

Good to look into why, probably just editor settings.

[kraftpunk97-zz]

Unable to identify issue. Although I did find the same behaviour here, if relevant.

[CarloLucibello]

cumsum has already been implemented. We can merge the cumprod part of this PR if you rebase and fix the formatting issues

[CarloLucibello]

this is no longer needed