Use nilearn masker memory parameters instead of safe_transform

[tsalo]

Summary

The nilearn masker objects have memory and memory_level parameters. I don't have much experience with them, but I think they could be used as an alternative to our safe_transform function (added in #520).

Additional details

Hopefully, this would reduce NiMARE's burden and simplify our code. I'm also hoping that nilearn's approach would handle large arrays more easily than our own.

Next steps

1. On a branch, replace safe_transform with simple Masker.transform calls, but initialize or set params in the maskers to use the memory parameters.
2. Memory profile our current code against the branch's approach.

[tsalo]

Since #681 was merged, do we even still need _safe_transform?

[JulioAPeraza]

I believe that after we merge this #752, CorrelationDistributionDecoder will be the only function that uses _safe_transform.
I will profile the memory of that function to see if it's worth keeping _safe_transform.