Implement Hann window classifier to reduce edge artifacts

[tayden]

This is a backwards-compatibility breaking change that remove the ability to specify batch_size and padding when doing classification but instead implements Hann Window stitching (see 1) to improve the quality of the processed outputs.

Basically, tiles are processed such that they overlap 50%, and then overlapping predictions are weighted and combined based on proximity to the center of the tile. Predictions for pixels near the center of the cropped section are assumed to have higher quality than pixel predictions near the borders. Thus, more central pixels are given higher weight when determining the final class and combining overlapping tile predictions.

There is also a fair amount of refactoring done in this PR, almost entirely to non-public APIs.

[tayden]

TO DO

• Update tests
• Update docs
• Show improvement using outputs in this PR

[tayden]

Image classified with moving window with size 512px, with 256px of context padding on each side:

Image classified with moving Hann window of size 1024px, where each overlaps 50% (comparable to above):

Larger crop sizes are still useful and may give better results (Here with size 3200px Hann window):