Consider creating patches on the fly

[rwood-97]

At the moment, we download maps, then patchify and save each patch as a png.
This essentially means twice as much data, i.e. if we have 1TB of maps, then we will also have 1TB of patches so 2TB total.

An alternative approach would be to patchify and save the pixel bounds for each patch but not the patch image itself. Instead, we could then create the patch on the fly by loading and cropping the parent image or by loading only the desired pixels of the parent image. This would minimise the storage needs when running mapreader.

[st7ma784]

Taken. Code to achieve this added to PR #545

• New Dataset class to override how patches are retrieved, given a patch df.