After 4bb1977b78 switched RCLASS_PRIME_CLASSEXT_WRITABLE_P to FL_TEST_RAW,
GCC lost the RB_FL_ABLE-derived hint that klass is not a special const,
and started warning that RCLASS_EXT_PRIME(klass) (used by the inlined
RCLASS_WRITE_SUPER in class_associate_super) could yield a near-zero
pointer. Make the existing VM_ASSERT(klass != 0) precondition visible
to the optimizer via RBIMPL_ASSUME so the warning goes away without
re-introducing the FL_TEST overhead.
Co-authored-by: Nobuyoshi Nakada <nobu@ruby-lang.org>
