🐛 `FSLMerge`: `OSError: [Errno 7] Argument list too long: '/bin/sh'`

Describe the bug

I believe the $ARG_MAX limit varies by system, but with long timeseries, where we split into individual timepoints in https://github.com/FCP-INDI/C-PAC/blob/de6407454d35e0cbf0434f7bc34fe44b924e2819/CPAC/registration/registration.py#L3844-L3847 and merge them back into a 4D image in https://github.com/FCP-INDI/C-PAC/blob/de6407454d35e0cbf0434f7bc34fe44b924e2819/CPAC/registration/registration.py#L3935-L3939, we can hit OSError: [Errno 7] Argument list too long: '/bin/sh'

Forcing relative paths via https://github.com/FCP-INDI/C-PAC/blob/2a9750c1a0075e5ba3f50ef3f54686d9357591b5/CPAC/utils/interfaces/fsl.py#L18-L25 was good enough to resolve the issue with the data I tested on brainlife.io, but users with ~2000 TR data are hitting the issue on bridges2

To reproduce

Run a config with

registration_workflows:
  functional_registration:
    func_registration_to_template:
      run: On
      apply_transform:
        using: abcd

with data with ~2000 timepoints

Preconfig

[X] abcd-options

Custom pipeline configuration

registration_workflows:
  functional_registration:
    func_registration_to_template:
      run: On
      apply_transform:
        using: abcd

Expected behavior

No crash

Acceptance criteria

[ ] C-PAC can recombine long timeseries into a 4D file regardless of system $ARG_MAX

C-PAC version

v1.8.5

Container platform

Singularity

Additional context

A few ideas of how to tackle this, combining capital letters with capital Roman numerals:

A. Check against `$ARG_MAX`

If the command length < $ARG_MAX, proceed as we have been. Otherwise, do one of these other things.

B. Do other thing regardless of `$ARG_MAX`

Just have one way of merging that is robust to OS command length limitations and always do this method.

I. Further customize `FSLMerge`

Ia. with `copyfile=False`

as suggested here

Ib. `cd` into `mapflow` directory

The relative paths all go up a level and down into a nibling mapflow directory, like ../applywarp_func_to_standard_222_/mapflow/. If we cd into that directory, we'd save ~33 characters per timepoint (I think Ia would save the same)

II. Use a Python merge utility

IIa. `nibabel.funcs.concat_images`

as suggested here and here

IIb. `nilearn.image.concat_imgs`

as suggested here

I'm sure there are other ways around this too

Where we’re hitting the error, it’s in the abcd method of applying the functional registration template. Looking at the code, I think the single_step_resampling_from_stc and dcan_nhp methods will also hit that error, but I think using this attached config could work ― the default method cuts the 4D files into 10-second 4D files to transform, while the other methods cut the 4D files into 3D files, so there’ll be 1/10 as many files to recombine with the default method, and then everything else will be the same as in abcd-options
FROM: abcd-options
pipeline_setup:
  pipeline_name: cpac_abcd-options_default-func-transform
registration_workflows:
  functional_registration:
    func_registration_to_template:
      run: On
      apply_transform:
        using: default

― cmi-cnl. Slack #user_support:thread:

FCP-INDI / C-PAC