FIX: Correctly cast nan when testing array_to_file- fixes ARM64 builds

[effigies]

Trying to replicate #861.

Closes #861.

[codecov]

Codecov Report

Merging #862 into maint/3.1.x will decrease coverage by 0.04%.
The diff coverage is n/a.

@@               Coverage Diff               @@
##           maint/3.1.x     #862      +/-   ##
===============================================
- Coverage        91.81%   91.77%   -0.05%     
===============================================
  Files               97       97              
  Lines            12359    12359              
  Branches          2177     2177              
===============================================
- Hits             11348    11343       -5     
- Misses             678      681       +3     
- Partials           333      335       +2     

Impacted Files	Coverage Δ
nibabel/environment.py	75.00% <0.00%> (-20.00%)	⬇️
nibabel/casting.py	85.28% <0.00%> (-0.87%)	⬇️
nibabel/dft.py	80.36% <0.00%> (-0.62%)	⬇️
nibabel/nifti1.py	92.12% <0.00%> (+0.15%)	⬆️
nibabel/volumeutils.py	84.52% <0.00%> (+0.38%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 97d5d03...68579bf. Read the comment docs.

[effigies]

According to https://docs.travis-ci.com/user/multi-cpu-architectures/#testing-on-multiple-cpu-architectures, we can also check on ppc64le and s390x. Will look into those if this works smoothly.

Also, there's OSX support, so we should at least test on that.

[effigies]

@matthew-brett The ARM64 jobs require building numpy from source every time, and we don't currently have the necessary C dependencies installed. Would it be reasonable to open a PR against numpy-wheels to add non-x86 wheels to their .travis.yml? Would anything else be required apart from updating multibuild?

[kaczmarj]

can you use miniconda? there are arm64 and powerpc miniconda installations, and numpy is compiled for arm64 and powerpc (https://anaconda.org/conda-forge/numpy).

[effigies]

Numpy is working on wheels on their end. I don't think it's too urgent to wait on that.

[effigies]

Anybody up for a review? Will merge Friday if nobody shouts.

[matthew-brett]

Looks right to me - thanks for tracking that down.

[effigies]

Great, thanks for having a look.