Make sure manylinux and macOS builds strip the binary

[mattip]

Once upon a time, we used more of multibuild in the openblas build process. That would pull in the STRIP_FLAGS from configure_build.sh. At some point that got left behind. I am not sure we are using the correct link arguments to strip the binaries, resulting in shipping uneccessary symbols. We should revisit adding -Wl,-strip-all to the CFLAGS in tools/build_steps.sh. As mentioned in numpy/numpy#29928, the windows was handled in #175.

This will reduce the size of the binary from ~24MB to ~23MB.

[mattip]

I am not sure the difference will change the size of the downloaded wheel, since the shared object compresses nicely.

[mattip]

We should revisit adding -Wl,-strip-all to the CFLAGS in tools/build_steps.sh

This breaks the tests, so a separate strip --all call is needed on the shared object.