native thread signal mask could block safe_read_tls_magic handling

One bug that occurred to me while analyzing #2335 is: the mask for a
native thread could have SIGSEGV blocked, meaning the safe_read_tls_magic
fault could be unhandled.

Of course SIGUSR2 could also be blocked, which would thwart taking over.
Our attach model relies on no non-pthreads threads at attach time (ok to
make them once we have control) (this is #2111) and on SIGUSR2 not being
blocked. We can't do too much about that, but we can ensure SIGSEGV is not
blocked if we're willing to spend a syscall every time at the top of
master_signal_handler_C().