Implement spatial rechunking in the dataset

While Pastis produces time-series of patches, the work in the internship is pixel-wise, and as been trained with large batches of pixels coming from different Pastis patches.

In order to do so, we would need to be able to rechunk the pastis patches into spatially smaller patches (maybe not as far as 1x1 pixels patches). Then we could increase the batch size while lowering the spatial size of patches, leading to more diversity within the same batch (at the expense of reading more patches from disk for a given batch).