Improve performance for large datasets and switch to multi-threading by default

[astrofrog]

This is a work in progress to address serious performance issues when reprojecting large datasets, mainly with reproject_interp.

To do:

• [x] Add more robust checks for hdu_to_vanilla_memmap to make sure we handle the case where it might fail (e.g. for an in-memory HDU)
• [x] Document TMPDIR and intermediate_memmap
• [ ] Write up recommendations for if someone wants to run using dask distributed with multi-processing

Fixes #393 Fixes #394

[codecov[bot]]

Codecov Report

Attention: Patch coverage is 81.21547% with 34 lines in your changes missing coverage. Please review.

Project coverage is 91.57%. Comparing base (38b134b) to head (7f7f90b). Report is 3 commits behind head on main.

Files	Patch %	Lines
reproject/mosaicking/coadd.py	74.78%	29 Missing :warning:
reproject/array_utils.py	91.07%	5 Missing :warning:

Additional details and impacted files

```diff @@ Coverage Diff @@ ## main #443 +/- ## ========================================== - Coverage 93.54% 91.57% -1.97% ========================================== Files 25 25 Lines 929 1021 +92 ========================================== + Hits 869 935 +66 - Misses 60 86 +26 ```

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.

[astrofrog]

Running into memory issues when mosaicking large datasets after the first one is done which suggests maybe an issue with the data getting loaded at the point where it is added to the main array.

[astrofrog]

Some CI failures related to zarr 3.0.0a1 released yesterday, need to investigate.

[astrofrog]

I'll go ahead and merge this and will then check the multi-processing and document any steps needed in a separate PR.