Inference.py: implement resampling of input raster

[remtav]

For certain applications like building footprints, it helps to upsample an input raster to a smaller spatial resolution (ex.: 50 cm to 25 cm) in order to get a finer extraction.

This will be implemented in the coming data_to_tiles.py, but will also be needed during inference to produce similar results on new images.

[ymoisan]

Seems to me the phrase "will also be needed during inference" comes up in several tickets. I think we are close to the point of viewing the inference task as more than providing a runtime for inference.py. What we really want I believe is a sound model deployment infrastructure, as per https://pytorch.org/tutorials/intermediate/flask_rest_api_tutorial.html.