Cleanup PDL::IO::FITS tests, fix incorrect tests; fix regression in writing 0D ndarrays

[djerius]

t/fits.t reorganization and fixes:

• many tests shared variables, requiring explicit re-initialization
• some regression tests for legacy FITS header writing were incorrectly excluded if Astro::FITS::HEADER was installed
• the test for tilde expansion in file names was never run.
• temporary files were explicitly unlinked instead of using the automatic facilities of File::Temp

Tests are now isolated into subtests, with no shared variables. The other issues were dealt with, but the level of header checking with and without Astro::FITS::Header is still quite asymmetric.

bug fixes:

• null hdus were always written using the legacy Perl code due to an incorrect check if Astro::FITS::Header was available.

regression fixes:

• passing 0D ndarrays to wfits caused an internal crash, as the code
  expected they be at least 1D. This used to work.

[coveralls]

coverage: 67.509% (+0.1%) from 67.399%
when pulling d9a7413 on djerius:pdl_io_fits
into 638ca16 on PDLPorters:master.

[mohawk2]

Please make the changes noted. It's a bit surprising that you submitted this without (from the CI results) even running it locally.

[mohawk2]

regression fixes:

• passing 0D ndarrays to wfits caused an internal crash, as the code expected they be at least 1D. This used to work.

Can you confirm in which version that used to work? It will need noting in Changes (which I'll update).

[djerius]

Please make the changes noted. It's a bit surprising that you submitted this without (from the CI results) even running it locally.

Harsh, dude. No need for that.

[djerius]

regression fixes:

• passing 0D ndarrays to wfits caused an internal crash, as the code expected they be at least 1D. This used to work.

Can you confirm in which version that used to work? It will need noting in Changes (which I'll update).

I know with certainty it worked back in 2.018. I'll try to track down the version where it changed.

[djerius]

regression fixes:

• passing 0D ndarrays to wfits caused an internal crash, as the code expected they be at least 1D. This used to work.

Can you confirm in which version that used to work? It will need noting in Changes (which I'll update).

I know with certainty it worked back in 2.018. I'll try to track down the version where it changed.

git bisect says 899a615.

git bisect run script:

git reset --hard
perl Makefile.PL
make -j80 && perl -Mblib -MPDL -MPDL::IO::FITS=wfits -E 'wfits { a => pdl(3)}, q{/tmp/foo}'