Tiling deskew

[pr4deepr]

This PR will deal with tiling large datasets.
Addresses #105.

When deskewing data in GPU, the deskewed image is converted to a 32-bit image. This results in increased GPU memory usage. This could potentially crash the software if:

• images are large in size
• GPU memory is small

To get around this, dividing the image into tiles and processing each image would be the best strategy. We already have a function for deskewing cropped images, so this can be reused here.

Strategy

• Estimate max allocated image memory for GPU (max_gpu_mem).
• Tiling strategy:
  • Get shape and size of a single deskewed image output
  • Create a dask image with same size and dtype (float32)
  • Divide into tiles, where each tile is less than max_gpu_mem. May add some padding of 5 - 10 px. Use dask rechunk and overlap?
  • Get the bounding box coordinates of each block (index of blocks and coords)
  • use crop_vol_deskew function to deskew each block where we pass the original image and roi coordinates in deskewed image space. Added bonus is that it accounts for deconvolution as well
  • Pass the cropped output into corresponding coordinates in the dask image.