The attached patch series refactors fragments from pthread_arch.h to
reduce redundancy and allow pthread_arch.h to be included at the top
of pthread_impl.h, where it can be used to influence the
definition/layout of struct __pthread, rather than the middle of the
file. This makes it possible to get rid of the duplicate canary and
dtv members that existed just to match multiple ABIs silumtaneously.

This involved a good deal of manual conversion/deduplication so it's
possible there are bugs for some archs. That's why I've posted the
series for review rather than just pushing. Reports of success/failure
(disassembly of pthread_self.o before/after probably suffice to
confirm no regression) would be very helpful.