Optimize compute_kda_ma for memory and speed

[adelavega]

• Implement "summary_array" return type for MKDAKernel, which convolves kernels to coordinates in a 3D dense volume, summing counts in place, saving substantial memory and compute.
• Using numba to speed up sphere kernel convolution
• @jdkent set types to int to reduce memory usage
• Minor improvements throughout

For large-scale MKDAChi2 (i.e. using Neurosynth dataset), memory footprint is reduced ~18-20x (25GB to 1.2GB), and computation is sped up ~3.2x.

[codecov]

Codecov Report

Attention: 16 lines in your changes are missing coverage. Please review.

Comparison is base (efae75e) 88.48% compared to head (a89c16d) 88.22%.
Report is 1 commits behind head on main.

❗ Current head a89c16d differs from pull request most recent head 6236f6c. Consider uploading reports for the commit 6236f6c to get more accurate results

Files	Patch %	Lines
nimare/meta/utils.py	67.44%	14 Missing ⚠️
nimare/meta/kernel.py	87.50%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #857      +/-   ##
==========================================
- Coverage   88.48%   88.22%   -0.26%     
==========================================
  Files          48       48              
  Lines        6337     6342       +5     
==========================================
- Hits         5607     5595      -12     
- Misses        730      747      +17     

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

[adelavega]

@jdkent these improvements make a small difference to run speed, but ultimately it's a pretty hard thing to optimize.

ultimately, it's just slow and fairly expensive process

[adelavega]

Even more speed increases by indexing the ijks array within numba

[jdkent]

LGTM! it's added and tested. @JulioAPeraza will be adding results for additional confirmation the results looks the same using the updated method.

[adelavega]

Awesome. Let's merge once everything passes, and cut a release soon.

[JulioAPeraza]

@adelavega, the new changes are fantastic! In the past, I could not train the decoder with more than 16 cores.

I used 40 cores this time, and it took 11 minutes to train an LDA-based decoder with 200 topics, and it only required 1.7 GB of memory.

@jdkent, I compared the decoders' results, and the output maps' values are exactly the same.

I also tested the new changes with different dataset sizes based on Neurosynth. See below:

[adelavega]

Julio, thank for your running this comparison. This is fantastic :)
This should make it much easier for people to replicate the results of your new paper.