croak(). Instead, the client is dropped, and an error is logged. If
we get too many errors (10), we bail. But, if we go a while without
getting any errors (60 seconds), we reset the error count. This is
so that if we have any bugs (or the node is just totally fubared),
we'll exit, but long-running daemons won't just accumulate the error
counter forever.

Also, handle a rare (but, on plab, who knows!) DOS-type condition with
blocking accept() noted in Stevens.