E.g. ones that throw interpreter exceptions, as those are currently
known to potentially cause leaks. Fixing the underlying leaks involves
the larger task of more IntrusivePtr usage.
Reference cycles may also cause leaks.
Reference cycles shouldn't occur but there's nothing really preventing
people from creating them, so may just as well be safe and deal with
them when cloning values. While the code is a bit more cumbersome this
way, it could actually be bit faster as well as it no longer caches
non-mutable values. (I measured it with the test suite: That's about
the same in execution time, maybe tiny little bit faster now;
definitly not slower).