Add 2D `conv`

It’s tricky to do efficiently so we should have an implementation here (even though it doesn’t need a new atom). We should also speed up the sparse matrix creation in the 1D convolution.

I implemented both here: https://discourse.julialang.org/t/convex-jl-slow-in-creating-constraints/34775/31

We could also let users reuse the matrix representation as is done there.