Implement parallelization across broadcasted dimensions

[astrofrog]

This builds on #374 to add a possible mode where if we are dealing with arrays with extra/broadcasted dimensions, we can split up the input array into chunks along the broadcasted dimensions which means we don't have to use the workaround of writing the whole dask array to a memmap.

For example, if we are reprojecting an array with dimensions (100, 2048, 2048) to (100, 1024, 1024), reprojecting only the celestial part of the WCS (the second and third dimensions of the array), we can specify e.g. block_size=(5, 1024, 1024) and this will cause the input to also be split up into chunks of (5, 2048, 2048) and each chunk of length 5 along the extra dimension will be handled separately.

One thing I'm not sure about is that at the moment we rely on the block size to fulfill certain conditions to toggle between the two main reprojection modes - but would it be cleaner to have a kwarg to do this?

@svank - your input would be appreciated too since you added the broadcasting functionality!

[codecov]

Codecov Report

Merging #376 (b07342b) into main (eabeb23) will decrease coverage by 0.48%.
The diff coverage is 87.50%.

@@            Coverage Diff             @@
##             main     #376      +/-   ##
==========================================
- Coverage   93.60%   93.12%   -0.48%     
==========================================
  Files          25       25              
  Lines         891      902      +11     
==========================================
+ Hits          834      840       +6     
- Misses         57       62       +5     

Files Changed	Coverage Δ
reproject/common.py	89.56% <87.50%> (-3.71%)	⬇️

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more