pin return type from downwards inference

[jmesserly]

along the lines of #27917, it would be nice if downwards inference required the return type to be exactly the context type, rather than a strong type.

I attempted to implement this (it is a trivial one line change after https://codereview.chromium.org/2456803004 lands) however it causes inference failures in some packages like stack_trace and unittest, not to mention in DDC and Analyzer code bases themselves. Because of that, it may be best to roll out independently of the above patch.

[jmesserly]

One common issue: the return type is weaker than the type parameter bound. That's a guaranteed inference failure with this change. We had some in our own code, for example T extends ParameterizedType but we assign the resulting T to a DartType variable.

[jmesserly]

My 1-line approach above doesn't work. I'm going to explore a different approach. We'll remember the type variables we solved for in the downwards phase, that did not contain ?, and make sure those are not inferred at all for the upwards phase.