otherwise it would result in an uncatchable exception and force the
program to terminate.  some exceptions are still considered super rare
and fatal, and they are still propagated (mainly from the io_service
and deadline timer).  errors on close() are just ignored (there
wouldn't be much we can do to recover in this case anyway).

Further, changed the handling of errors on the result of async_accept():
instead of stop accepting on some errors, log them and keep accepting
new connections.  on a closer look at the code I believe that's more
appropriate; silent stop would be rather confusing (if the error were
really fatal we should rather invoke an exception and trigger termination),
and cases like EMFILE definitely shouldn't cause this situation.
errors that can caused by stop() should still result in stopping accept(),
and the new code behaves that way.