RF: Deprecate longdouble<->int hacks

[effigies]

We had a pair of functions as_int() and int_to_float() that handled bugs in numpy that existed at least as recently as 1.11, but which no longer exist in 1.20, our current minimum.

This PR replaces as_int(...) with int(...) and int_to_float(..., dtype) with dtype(...) and deprecates these functions. Regression tests to ensure that these native Python/numpy functions work as expected are retained.

[effigies]

@matthew-brett You added these back in #70, but I think I'm right that they can go away at this point. Any objections? Any other regression tests worth putting in?

@larsoner or @mscheltienne Would you mind having a review, since you were just dealing with these?

[codecov]

Codecov Report

Attention: 1 lines in your changes are missing coverage. Please review.

Comparison is base (dabfc46) 92.25% compared to head (bbfd009) 92.21%.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1272      +/-   ##
==========================================
- Coverage   92.25%   92.21%   -0.04%     
==========================================
  Files          99       99              
  Lines       12466    12455      -11     
  Branches     2564     2561       -3     
==========================================
- Hits        11500    11486      -14     
- Misses        643      645       +2     
- Partials      323      324       +1     

Files	Coverage Δ
nibabel/casting.py	86.17% <100.00%> (-1.06%)	⬇️
nibabel/conftest.py	93.75% <100.00%> (-6.25%)	⬇️
nibabel/arraywriters.py	96.86% <91.66%> (ø)

... and 1 file with indirect coverage changes

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[larsoner]

Much simpler!

[effigies]

Thanks for the review!