Add flag to do test time augmentation to improve prediction robustness

[tayden]

This PR adds a boolean flag to the CLI and lib functions which enables test-time augmentation in the find-kelp and find-mussels functions.

For each image chip, it is classified in it's normal orientation, then repeatedly rotated 90 degrees and reclassified. The chips are then flipped horizontally, reclassified, and finally flipped vertically and reclassified. In total, each chip is classified 6 times in different orientations. The predictions are then un-flipped and rotated such that each chip has 6 predictions in the correct orientation. Finally, individual pixels are classified using a combination of all these predictions such that the most confident prediction over all the possible orientations is used as the final output.

The result of this is that processing an image takes about 6 times as long as in the normal, non-tta, mode to process. However, the outputs are generally expected to be of higher quality.