llvm::sys::ExecuteAndWait timeouts break when called from multiple threads

On Unix, LLVM's ExecuteAndWait/Wait implement SecondsToWait with alarm() and a SIGALRM handler. In a test runner that launches child processes from several worker threads, the alarm is process-wide and the signal can land on another thread, so waitpid is never interrupted and a hung child blocks forever. Fix: start the child with llvm::sys::ExecuteNoWait, then loop llvm::sys::Wait(process, 0, &err, nullptr, /*Polling=*/true) with a short sleep, checking a per-test steady_clock deadline; on expiry kill it yourself (kill(pid, SIGKILL) on Unix, TerminateProcess(process.Process, 1) on Windows) and reap with a blocking Wait. Also note ExecuteAndWait returns -2 for both crashes and timeouts, so you can't tell them apart from the return code alone.