Implement several speedups for polar view generation

[psavery]

Most of these speedups are applicable for generating a whole bunch of polar view in a row with the same image (i. e., cache_coordinate_map=True when instantiating the PolarView class), but a few of them also improve performance of creating the PolarView in general.

Overall, we achieved at least a 10x speedup (in serial) for generating a bunch of images in a row, and even more of a speedup for some examples. A list of the improvements are as follows:

1. Using a "moving sum" for summing the detector images. This is particularly valuable for when you have many detectors (i. e., 32 subpanel Eiger). It eliminated the requirement of stacking all the images together before performing the sum, and we can instead discard images immediately after they are included in the sum.
2. Precompute bilinear interpolation parameters, including the weights and indices of detector pixels used for each polar pixel. As long as the instrument doesn't change, these interpolation parameters don't change either, and thus we don't need to generate them every time. They are now included in the cache map.
3. Run bilinear interpolation in numba. This is faster in the single-threaded case, and it also improves parallelism via threading since numba releases the GIL.
4. Instantiate and provide a re-usable output buffer for the interpolation. We re-use the output buffer for every detector, so that we don't have to repeatedly instantiate a chunk of memory for the output for every detector, as we were doing before. - superseded by number 6.
5. Generate nan mask ahead of time. When we reuse the same instrument for many images in a row, we can now reuse the same nan mask, since it should not change. Skipping the step where we regenerate the nan mask provided a considerable performance improvement.
6. Avoid creating the full-size intermediate image immediately before the moving sum. Instead, we can just add valid values to the moving sum directly. This provided yet another big speedup.
7. Completely eliminate all intermediary arrays, so that the weighted interpolation is always computed and set directly into the output array.

I tested these changes a fair amount, and the output looks completely identical to what it was before.

[codecov]

Codecov Report

❌ Patch coverage is 86.88525% with 8 lines in your changes missing coverage. Please review.
✅ Project coverage is 49.29%. Comparing base (06a6678) to head (f88a57d).
⚠️ Report is 13 commits behind head on master.

Files with missing lines	Patch %	Lines
hexrd/instrument/detector.py	72.00%	7 Missing ⚠️
hexrd/projections/polar.py	97.22%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #833      +/-   ##
==========================================
+ Coverage   49.26%   49.29%   +0.03%     
==========================================
  Files         143      143              
  Lines       23109    23135      +26     
==========================================
+ Hits        11385    11405      +20     
- Misses      11724    11730       +6     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[psavery]

Chris Budrow tested and approved these changes.