Allow generating confounds at resampling stage

[effigies]

What would you like to see added in fMRIPrep?

Or possibly make it its own stage?

Do you have any interest in helping implement the feature?

Yes

Additional information / screenshots

No response

[effigies]

Related to call w/ @FeHoff, @asmacdo, @jbwexler, @yarikoptic:

We should split out confounds time series into trivial (already present / arithmetically derivable from timeseries present in fit workflow, e.g., motion, FD), cheap (resample BOLD to native, derive, e.g., dvars) and expensive (e.g., compcor).

--level resampling should emit trivial and cheap. --level full should continue to output the expensive.

Open questions:

1. Should --level minimal emit trivial confounds?
2. How should we group these? (desc-motion, desc-rois, desc-compcor, ...)?

As written, this is a breaking change. Could consider back-porting to LTS, and enable with a flag.

[FeHoff]

1. I would vote for leaving --level minimal as is to keep changes small and emit cheaper confound time-series in --level resampling (all but comp_cor, edge_comp ?)
2. Regarding grouping of confound measures, the classical *desc-confounds_timeseries.tsv could stay and compcor and associated time-series could get an own *desc-confounds+compcor_timeseries.tsv