* BUILD: update OpenBLAS to 0.3.21 and clean up openblas download test

* set LDFLAGS on windows64 like the openblaslib build does

* use rtools compilers on windows when building wheels

* fix typos

* add rtools gfortran to PATH

* use the openblas dll from the zip archive without rewrapping

* typos

* copy dll import library for 64-bit interfaces

* revert many of the changes to azure-steps-windows.yaml, copy openblas better in wheels

* fix wildcard copy

* test OpenBLAS build worked with threadpoolctl

* typos

* install threadpoolctl where needed, use for loop to recursively copy

* update macos OpenBLAS suffixes for newer gfortran hashes

* use libgfortran5.dylib on macos

* fix scripts

* re-use gfortran install from MacPython/gfortran-install on macos

* use pre-release version of delocate

* fixes for wheel builds/tests

* add debugging cruft for pypy+win, macos wheels

* add DYLD_LIBRARY_PATH on macosx-x86_64

* use 32-bit openblas interfaces for ppc64le tests

* skip large_archive test that sometimes segfaults on PyPy+windows

Not sure what the usual timeline on updating this is.
Running the CI on this PR is one way to narrow down the Cygwin problems addressed in PR#20654.
If this build also has four failures, then OpenBLAS is probably the problem.
If this build uses the new OpenBLAS with no problems, then I need to see what else changed
in the Cygwin packages around the time the builds started failing.

When working with ILP64 in Windows the correct file is downloaded, but
is always copied to "openblas.a". This breaks distutils that looks for
"openblas64_.a"