Lazy Patches

[d-chambers]

We need 'em.

Essentially, each patch that doesn't already have the data numpy array loaded into memory would lazily load the data when the data attribute is first accessed. This will probably be accomplished with fsspec and will make it easy to apply map-type parallelism to spools using Executors/Clients from various packages.

[d-chambers]

We might also consider a load method to ensure all the data are loaded, eg for debugging parsing. We will need nice error messages as not to overly confuse users if a parsing error crops up during a processing step due to the Patch's lazy nature.

[d-chambers]

I am thinking we wont do this; just having an option to split spools would effectively be the same things while keeping the code more simple.