handler's stack doesn't stomp siginfo. this also fixes !__HAVE_SIGINFO, in that case pthread__signal_tramp assumes uc->uc_stack.ss_sp points the old signal mask. pointed by uwe@.