Padding strategy for tiles

[forklady42]

Training on original grid sized data triggered CUDA OOM errors. Our most promising path to significantly reducing CUDA memory is a tiling approach where a section of the crystal structure is used each time.

These tiles will not necessarily line up with the periodic boundaries of the crystalline structure, which makes using circular padding for the convolutions incorrect. This is a place to write up various tests of different padding strategies, e.g. zeros padding or slicing a halo around the tile, in conjunction with tiling.