Reimplement SCS in terms of conv2d

[jatentaki]

This PR proposes a change in implementation of SCS which uses Conv2D, a much more performant primitive. The gist of the idea is to

1. Normalize the kernel as we already do
2. Compute per-window normalization factor coming from the image by sqrt(avg_pool2d(input ** 2) * window_size)
3. Compute the dot products via conv2d between the input and normalized kernel
4. Normalize the output elementwise by the array obtained in step 2.
5. Proceed as before

On my laptop this achieves more than 2x performance improvement. I attach a quick test (compat_test.py) to verify that the two implementations yield (almost) the same results. This PR will obviously need some code unification and renaming before being merged, I submit it as it to facilitate easy comparisons by the maintainers :)

[brohrer]

This is fantastic. I look forward to doing a careful comparison of results and a clean integration. I hope I can get to it sooner rather than later. Thanks!

[enzokro]

This implementation is great! Thank you for making it.

I have a quick question about eps, but it shouldn't be a difference-maker.
For the sake of keeping its scale, should we add it when squaring the input:

       x_norm_squared = F.avg_pool2d(
            (x + self.eps) ** 2,
            kernel_size=self.kernel_size,
            stride=self.stride,
            padding=self.padding,
            divisor_override=1, # we actually want sum_pool
        ).sum(dim=1, keepdim=True)

So when the outputs are normalized below:

    y = y_denorm / (x_norm_squared.sqrt() + q_sqr)

It will be with eps instead of sqrt(eps)?

[jatentaki]

@enzokro you're right, it would make more sense to add it before avg_pool2d. I somehow missed that this addition can be moved there in order to completely preserve the original semantics.

[brohrer]

Thanks for this @jatentaki ! It looks like several contributors had similar ideas at the same time. A similar innovation from @ClashLuke has been folded into the main branch. I'll go ahead and close this PR for now, but thank you very much for spending time thinking about this.