ENH: Split SyN fieldmap estimates per-EPI

[effigies]

This is an experimental PR to undo the merging of individual EPI images which may not be well-aligned with respect to the phase-encoding direction. This will revert us to the basic collation method of fMRIPrep 20.2.x while preserving the fit-apply split introduced in fMRIPrep 21.0.0.

This experiment is worth performing before continuing to optimize b-spline approximation of the SyN-estimated field.

I've tried to write this with an eye to making it easier to plug in grouping strategies in the future.

First step in fully addressing #306.

[effigies]

When splitting, I get normal looking fields (only one shown, as they all look alike):

But the conversion to the individual subject space varies across images (I see two like each):

Unsurprisingly distortion correction looks much more reasonable when the run-space fieldmaps look reasonable (though neither is as bad as some failures we've seen):

What's odd is that all the input images have the same affine and the only difference I'm seeing in metadata is a few of the slice times.

[effigies]

Oh man. There's a collation problem here, where the low-frequency maps are associated with odd runs and high-frequency with even.

[effigies]

Addressed in #317.

[effigies]

This does not seem to be necessary for the test cases I'm looking at. I suggest we hold off until 23.0 to see whether problems are resolved for most users with our other updates. Then we could make this an option.