Instead of separate locks for buffering and sender modes
configuration, use the same lock.

Now that disable_buffering and pre_fork may both try to stop the flush thread
at the same time, attempt to stop an already stopped thread does not
indicate a bug in the client code.

Single field assignement is atomic according to Python FAQ, so we
don't need to protect reads and writes to self._queue as such. The
only place where we need a lock is when stopping the sender, to ensure
that the Stop marker is the last thing to be placed in there.

https://docs.python.org/3/faq/library.html#what-kinds-of-global-value-mutation-are-thread-safe

os.register_at_fork is not available on all python versions, notably
python 2.7. While we check for the availability explicitly, mypy
cannot see through that and complains.

The method is now thread safe, and can be used with forking applications.

Re-format parameter descriptions so they are rendered correctly in the
html docs.