Update TorchIO version

[fepegar]

Resolves #672.

The new version of TorchIO fixes a small bug discovered recently which affects patch-based inference:

• Fix overlapping in patches aggregator TorchIO-project/torchio#832

[javier-alvarez]

Can you clean the description of the PR?

[fepegar]

Done. Is that better?

[fepegar]

This happens because InnerEye is asking TorchIO to place a patch in a location where a larger patch is meant to be, i.e., the input size to the aggregator is smaller than the sampled patch. This happens when the model output is smaller than the input.

I have modified TorchIO so the error message when this happens is more informative.

I will submit a PR here to pad the model outputs when this situation occurs. When the model outputs are expected to be smaller than the input, a large overlap is typically specified anyway, so the patches will be cropped inside the aggregator later on, so it doesn't matter much how they are padded initially.

Hopefully, all that made some sense.