Host components outside the setState path are sometimes unnecessarily diffed and updated

[gaearon]

Extracted from #12643.

We're currently diff a host node in the complete phase even if it has bailed out in the begin phase. By itself, this just means we're doing a little bit more work than necessary. But it also has some observable effects. In particular, #13424 (which by itself would be very rare and probably not worth fixing) is made worse because we end up committing update to inputs from another component.

I have a fix for this in #13423. Filing this for posterity.