Failed or interrupted downloads leaves corrupted template files.

[bpinsard]

If a templateflow image download fails, the truncated file is non-zero and considered as fetched and returned in subsequent queries.

This occurred in a rare case where the truncated image silently propagated through Ants-based resampling and subsequent brain extraction, corrupting the whole pipeline.
nipreps/fmriprep#2192
It seems that it was using the default non-datalad download.

An option would be to checksum-validate each time the template is queried.

[oesteban]

Using the datalad backend this should not occur. I'm not sure it is worth the pain maintaining a MANIFEST file with MD5 sums. Perhaps we could reconsider making datalad the default since it is more stable now?

[oesteban]

Opening with nibabel could be an easy tradeoff between today and a decent validation of checksums.