Blacklist the "threaded-ml" thread, which seems to mask the
backtrace signal.

Improve signal-handler synchronozation, to avoid segfaulting when
giving up on waiting for all threads to handle the signal.
Furthermore, when one or more threads fail to handle the signal in
time, return a GimpBacktrace instance with backtraces for all the
other threads, and with empty backtraces for all the non-responding
threads, instead of returning NULL and leaking the allocated
instance.  Don't blacklist threads that failed to handle the signal
in time, and instead shorten the wait period for handling the
signal, and yield execution during waiting to lower the CPU usage.