Improve performance in _manylinux.py

[cthoyt]

This PR makes a few performance improvements:

1. Reduces the doubled call to _is_compatible(arch, glibc_version)
2. Moves the assignment of the tag variable directly into the yield, reducing memory allocation in case this is never used when _is_compatible(arch, glibc_version) is false
3. Uses an assignment expression for dictionary lookup