Make square centre-padding a model transform

This MR fixes issues #23 (closed) and #39 (closed), by:

• Removing centre-cropping from all databases raw-data loaders
• Creating a SquareCenterPad transform that operates on tensors
• Adding that transform to all relevant models (Pasa, Alexnet, Densenet)

src/mednet/engine/saliency/viewer.py

@@ -226,6 +226,7 @@ def run(
         for sample in tqdm(
             dataset_loader, desc="batches", leave=False, disable=None
         ):
+            # WARNING: following code assumes a batch size of 1. Will break if not the case.
             name = str(sample[1]["name"][0])
             label = int(sample[1]["label"].item())
             data = sample[0][0]
@@ -243,7 +244,9 @@ def run(
             # regions of interest.  We need to abstract from this to support more
             # datasets and other ways to annotate.
             if show_groundtruth:
-                ground_truth = sample[1].get("bounding_boxes", BoundingBoxes())
+                ground_truth = sample[1].get("bounding_boxes", BoundingBoxes())[
+                    0
+                ]
             else:
                 ground_truth = BoundingBoxes()