Feature: Implement KMeans attribution support and TakesMost LRP rule

[jackmcrider]

This commit introduces several new components to enhance k-means based model attribution and provide a new LRP rule for pooling layers:

1. New Layers (src/zennit/layer.py):

   • PairwiseCentroidDistance: Computes pairwise distances between inputs and a set of centroids.
   • NeuralizedKMeans: A layer representing k-means discriminants as a linear transformation (tensor-matrix product with bias).
   • MinPool1d and MinPool2d: Min-pooling layers for 1D and 2D inputs, implemented by negating inputs/outputs of MaxPool.

2. New Canonizer (src/zennit/canonizers.py):

   • KMeansCanonizer: Replaces PairwiseCentroidDistance layers (with power=2) with a sequence of NeuralizedKMeans, MinPool1d, and torch.nn.Flatten. This neuralization enables the application of LRP rules to k-means like clustering outputs.

3. New LRP Rule (src/zennit/rules.py):

   • TakesMost: An LRP rule designed for max-pooling layers (and by extension, min-pooling layers after neuralization).
   • Distributes relevance based on a softmax of input contributions within each pooling window, providing a "soft" alternative to winner-takes-all.
   • Implements window-wise maximum subtraction for robust numerical stability during the softmax calculation.
   • Dynamically adapts internal convolution, pooling, and transposed convolution operations to match the input dimensionality (1D/2D).
   • Utilizes common, refactored parameters for unpooling operations.
   • Respects the original module's hyperparameters.

4. New Tests (tests/):

   • test_kmeans_canonizer in test_canonizers.py: Verifies the correctness of the KMeansCanonizer, ensuring module replacement, functional equivalence of cluster assignments, and proper restoration.

Replaces #197
Closes #198

[jackmcrider]

Some tests fail, but I think these are mostly lint errors coming from the main branch.

[chr5tphr]

Hey Jacob,

Thanks a lot for finalizing the PR! I will get rid of the tests on Python 3.7, so that's fine.
The pylint errors are also fine, I do not think the number of positional arguments is too bad.
The flake8 is valid, you can just add a whitespace where reported.

Could you go through the places that are missing in the coverage report? Some might be unrelated to KMeans and TakesMost, so you can ignore those. It would be good to have those places at least called in some test.

 Name                       Stmts   Miss Branch BrPart  Cover   Missing
----------------------------------------------------------------------
src/zennit/canonizers.py     144      2     34      1    98%   381-382
src/zennit/composites.py     128      0     34      1    99%   159->161
src/zennit/layer.py           27      1      0      0    96%   157
src/zennit/rules.py          138     43     14      0    69%   455-460, 464, 468, 471, 475-482, 486-499, 503-507, 510-564
----------------------------------------------------------------------
TOTAL                       1123     46    312      2    96%