Average overlapping predictions in patch-based inference

[fepegar]

There are some quantitative differences in inference depending on whether RadIO or TorchIO are used:

• Replace RadIO with TorchIO for patch-based inference #666 (comment)

I suspect that the little differences might happen because overlapping patches are cropped by default in TorchIO, while they seem to be averaged in RadIO. TorchIO's aggregator has an averaging mode, so I'll modify the settings in this repo (after updating the TorchIO version) and see if that gets us the exact same results.

AB#5101