[ADD] Use TN formulation of Dangel, 2023 to compute average patches

[f-dangel]

Implements #60.

I added a test that ensures the TN formulation computes faster and uses less memory than the previous approach.
There are still more performance improvements possible for structured convolutions, but I first need to put more work into the einconv library's simplification mechanism before we can make those even faster and more memory-efficient.

Here is the output of the test case (CPU):

Memory used by inefficient function: 2537.1 MiB.
Memory used by efficient function: 1161.7 MiB.
Inefficient function took 1.09e+00 s
Efficient function took 5.12e-01 s

[runame]

LGTM. I think this is an exciting improvement.