[BUG][BLD] revert DEF component of #21878

[jbrockmendel]

#21878 introduced a subtle build problem that is not caught by the CI. Running python setup.py build_ext --inplace followed by python3 setup.py build_ext --inplace causes compile-time errors in py3 (or if running these in the opposite order, errors in py2). These are fixed by running python setup.py clean in between.

This reverts the relevant changes.

For anyone curious: cython supports syntax:

DEF foo = [...]

IF foo:
    [...]
ELSE:
    [...]

and these IF/ELSE conditions get evaluated at compile-time. #21878 incorrectly assumed that "compile-time" in this context meant ".c -> .so" time, not ".pyx -> .c" time. In this we used:

DEF PY2 = str == bytes

IF PY2:
    [...]
ELSE:
    [...]

so after running setup.py build_ext --inplace in py2, the ".c" file we end up with has already gotten rid of the PY3 branches. When we run python3 setup.py build_ext --inplace cython uses the existing .c file, tries to compile it to .so, and breaks.

[codecov]

Codecov Report

Merging #21919 into master will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master   #21919   +/-   ##
=======================================
  Coverage   91.96%   91.96%           
=======================================
  Files         166      166           
  Lines       50323    50323           
=======================================
  Hits        46281    46281           
  Misses       4042     4042

Flag	Coverage Δ
#multiple	90.36% <ø> (ø)	⬆️
#single	42.23% <ø> (ø)	⬆️

---

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 63eff04...d939bde. Read the comment docs.

[jbrockmendel]

@jreback @WillAyd quick take before I merge?

[jreback]

lgtm

yeah ideally these would be compile time checks