Introducing sparse atomic corrections to GPAW new

This MR does the following:

1. The corrections have been moved out into a function.
2. The dense atomic overlap corrections have been batched across atoms. For small systems (<~500 basis functions) where corrections can in principle be dense (sparse corrections lose maybe ~0.5 seconds in performance), the increase in cost due to the matmul over a block diagonal dense matrix is completely overshadowed by the savings due to not doing the repeated additions to S_MM (although both these numbers are tiny).
3. The sparse corrections are also batched across atoms similar to the old GPAW. For systems with tens of thousands of basis functions/gamma point, the savings can exceed 30-50x.

So why is this a draft?

I don't yet know what would be an accepted way of exposing the sparse keyword. I also don't know how to deal with the mypy error.

I don't know if we have band/orbital parallelism in the new lcao yet, in that case we would need to also keep track of the basis Mstart Mstop indices again.