Fixes#96319
The logic around handling co-inductive cycles in the evaluation cache
is confusing and error prone. Fortunately, a perf run showed that it
doesn't actually appear to improve performance, so we can simplify
this code (and eliminate a source of ICEs) by just skipping caching
the evaluation results for co-inductive cycle participants.
This commit makes no changes to any of the other logic around
co-inductive cycle handling. Thus, while this commit could
potentially expose latent bugs that were being hidden by
caching, it should not introduce any new bugs.
40c6d838cd