The comment explains it all, To avoid `zlib` blowing up the memory
usage, we need to reduce the chunk size a lot.

The downloads are now done in chunks, allowing an optional callback
to be provided which is checked between every chunk. The callback is
expected to return a boolean dictating whether to cancel the download
or not.

This drops the use of `multiprocessing.Process` objects to (down)load
volumes, which should fix the pickling troubles on macOS.

Additionally, the thread can now react to interruption requests and
gracefully cancels downloads or abruptly terminates loads.

This allows calling code to handle cases where the user cancelled the
(down)load.

This makes use of processes to load NumPy arrays of the volumes in a
separate environment, allowing it to be cancelled without freezing the
main GUI. Additionally, the normalisation can now be done "fast",
meaning that now floating, intermediary arrays are used, saving a bit
of speed and a lot of memory.