- documentation in numpydoc format
- pylint + flake8 stuff
- KMeansCanonizer
- NeuralizedKMeans layer
- LogMeanExpPool layer
- Distance layer
- Distance type

- Explaining Deep Cluster Assignments with Neuralized K-Means on Image Data
- I tried to adhere to guidelines
- That means: random data, random weights
- Code for real data and real weights in comments
- Runs on colab, did not test blender
- also adds the reference to docs/source/tutorial/index.rst

Co-authored-by: Christopher <15217558+chr5tphr@users.noreply.github.com>

change `torch.log(torch.tensor(n_dims, dtype=...))` to `math.log(n_dims)`

Co-authored-by: Christopher <15217558+chr5tphr@users.noreply.github.com>

change `setattr(parent_module, ...)` to `parent_module.add_module(...)`

Co-authored-by: Christopher <15217558+chr5tphr@users.noreply.github.com>

add spaces around binary operators

Co-authored-by: Christopher <15217558+chr5tphr@users.noreply.github.com>

- rename Distance to PairwiseCentroidDistance
- remove LogMeanExpPool (might become relevant again, but not for now)
- add MinPool1d and MinPool2d in layer.py
- add MinTakesMost1d, MaxTakesMost1d, MinTakesMost2d, MaxTakesMost2d rules
  - largely untested. especially kernel_size as int or kernel_size as tuple
  - in principle, MaxTakesMost2d should also work for MaxPoll2d layers in
    standard conv nets - but needs some testing
- add abstract TakesMostBase class
- remove type definition for Distance in types.py
- adapt KMeans canonizer:
  - replace LogMeanExpPool with MinPool1d followed by torch.nn.Flatten
  - remove beta parameter; beta is now sit in MinTakesMost1d
  - remove deepcopy and simply return the module itself
- update docs/src/tutorials/deep_kmeans.ipynb
- doc strings

- merge changes coming from github web interface

- various non-functional changes