Incremental flushing of messages in fine grained incremental mode

[ilevkivskyi]

I previously proposed to switch from alphabetic order of processing of targets to topologically sorted w.r.t. to dependencies. The argument was that it may give some minor performance boost because we will reach a fixed point faster. However, there is another argument, in this case we can incrementally flush errors, because after we fully processed all targets in an SCC there is no way back.

[gvanrossum]

But since fine-grained mode is usually invoked from the daemon we'll also need a re-architecting of the daemon's client-server protocol to allow returning errors as they are produced. @msullivan implemented that for non-daemon mode a while back. It's a great improvement over the original "flush at end" approach, and (for runs that take longer) I miss it.

[msullivan]

This is a dupe of #4702 but I'll close that one. I think this would work, but it will require a rework so that we don't depend on error reprocessing (#4801)