Harden GADT constraint handling to survive illegal F-bounds

[odersky]

This currently fails BestEffortCompilationTests

The problem is that i20317a.scala contains illegal cyclic references. The cycles cause many different algorithms in the compiler to loop. We normally reject the program, and that's that. But under best-effort, we crash with Stackoverflows in the loops later. What to do? It's not feasible to avoid the looping behavior at acceptable cost. Can we disable best effort under certain conditions?

[jchyb]

The fix is to put the file name into compiler/test/dotc/neg-best-effort-unpickling.blacklist. This will cause this specific test to not be executed and the CI to pass, as it does with other files test files related to cyclic reference errors. In the real world IDE scenario we let it crash with the message and handle that gracefully on an IDE side (I know this is far from a perfect solution, and there are a few other yet-unsolved cases like cyclic-reference-errors).

Usually in CI we should get a message asking to put the file in compiler/test/dotc/neg-best-effort-unpickling.blacklist. but I see it does not show there now (which in hindsight, makes sense), so this is definitely something I need to fix.

In general I would not want the compiler development to be throttled by the best-effort options in any way, so using the aforementioned exclusion list is very much expected. Again, the message with the exclusion list it was supposed to show, but it makes sense it wouldn't with the stack error.

I'll figure something out for cyclic-reference errors in the coming weeks for 3.5.1. I apologize for the inconvenience!