Option to run FreeSurfer in working directory instead of output

[effigies]

Running FreeSurfer on clusters seems to be prone to NFS sync failures, which is difficult to resolve, as FreeSurfer is run directly in the output directory. While some nipype sync issues may be resolved by placing the working directory on /tmp or some other local scratch that has higher consistency guarantees, to do the same for FreeSurfer would mean users then need to copy the final outputs to the output directory.

An alternative would be to allow a --fs-temporary-directory flag that would run recon-all on a local filesystem and copy to the output directory. I would recommend the default be <work_dir>/freesurfer, but allow users to specify another location. It should be noted that one effect is to result in two copies of the data, which may trip up users running locally.

Example issue (can dig up more if we need evidence):

• https://neurostars.org/t/fmriprep-autorecon31-error/3227

This is somewhat related to #27.

[oesteban]

I'd rather do it always on <work_dir>/freesurfer and then move everything out to <out_dir>/freesurfer. WDYT?

[oesteban]

BTW, this issue is somewhat related to #41. We could try to address both together

[oesteban]

And, unless you point the work directory to a CIFS filesystem, the proposed solution should also address #30.

[effigies]

The problem is that if somebody uses a prerun fs directory, then we copy it into the container, do nothing, and copy it back.