perf: transformer batching

[danellecline]

BREAKING CHANGE: Long-overdue performance improvement on clustering through batching. Also, some work related to #10 and the removal of unused imports and code. Clustering should run faster on both CPU/GPU. The default model is now google/vit-base-patch16-224 which performed better on UAV images. Requires an update to your config.ini. e.g.

# google/vit-base-patch16-224 is a model trained on ImageNet21k with 21k classes good for general detection
# dino models were pretrained on ImageNet which contains 1.3 M images with labels from 1000 classes
# Smaller block_size means more patches and more accurate fine-grained clustering on smaller objects
# Larger block_size means fewer patches and faster processing
model = google/vit-base-patch16-224
;model = facebook/dino-vits8
;model = facebook/dino-vits16