Better handling of sigset_t on LP32.

The main motivation here is that the sigprocmask in pthread_exit wasn't
actually blocking the real-time signals, and debuggerd (amongst other
things) is using them. I wasn't able to write a test that actually won
that race but I did write an equivalent one for posix_spawn.

This also fixes all the uses of sigset_t where the sigset_t isn't
exposed to the outside (which we can't easily fix because it would be
an ABI change).

Bug: https://issuetracker.google.com/72291624
Test: ran tests
Change-Id: Ib6eebebc5a7b0150079f1cb79593247917dcf750