Clarify and fix accuracy in eval public API, on reporting "unable to evaluate."
While the eval internals always used the flag pair valid and cfavalid almost correctly, the public interface exposed the outcome as a single flag, and the various interpretations of this flag were a mess.
Old cfacc treated sizeof(whatever) in some contexts as "known to be zero," which is wrong.
The generally correct answer, which new cfacc now uses is, "unknown now, but GCC will see it as a fine constant."
New tests/eval.cfa captures this fact: is runnable and would fail on old cfacc; it passes with new cfacc.
