cyclic_default_values_test stack overflows during resolution

[jmesserly]

Just caught this testing DDC against latest Analyzer. Seems likely to be caused by #25794 fix.

[jmesserly]

(I'm a bit surprised by a stack overflow though, as we don't do anything recursive. What I'd expect is that inference would just not work for one of the two functions. edit: that wouldn't be correct behavior either, but it would be a bit less problematic.)