Implement caching for Estimators and Transformers

[JulioAPeraza]

Closes #844.

Changes proposed in this pull request:

We are implementing caching at three levels of a meta-analysis:

• High-level (or estimator-level): caches _fit()
• Mid-level (or transformer-level): caches _transform()
• Low-level (or modeled activation (ma)-level): caches _get_ma_map()

The most typical use case will be at lower levels, given that it is very common to recompute the same MA maps when working with the same database (e.g., Neurosynth).

Transformer, at the mid-level, will benefit as well from caching.

At the estimator level, caching was implemented just for the sake of completeness. We currently recommend saving the MetaResult object to a pickle file and loading it again if will be reused. This reduces the large overhead that comes from hashing a whole NiMARE database object.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (720c5c6) 89.12% compared to head (391f52e) 89.13%.
Report is 1 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #845   +/-   ##
=======================================
  Coverage   89.12%   89.13%           
=======================================
  Files          49       49           
  Lines        6142     6155   +13     
=======================================
+ Hits         5474     5486   +12     
- Misses        668      669    +1     

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

[JulioAPeraza]

This shows the performance at memory level 1 (caching only _fit()). With memory level 2 (caching _fit() and _transform()) the performance was the same. However, at memory level 3 (caching _fit(), _transform(), and _get_ma_map()) I saw a minimal decrease in the performance but still better than no caching.

[JulioAPeraza]

I tested the hierarchy inverted:

Memory level 1: caching _get_ma_map().
Memory level 2: caching _get_ma_map(), _transform().
Memory level 3: caching _get_ma_map(), _transform(), and _fit().

However, the performance at level 1 with caching was worse than no caching, probably because the computation in _get_ma_map() is not that expensive.

[jdkent]

LGTM!