deriving codegen: fold comparisons of variants without substructure into the catch-all

[pnkfelix]

Spawned off of #15503 (comment)

I was thinking of the particular case of C-like enums on the way home today.

Those cases are covered by this (not yet implemented) optimization: I really should be letting comparisons of two enums that both have no subparts (e.g. B and D in enum E { A(int), B, C(f64), D } fall through the catch-all as well, since the catch-all will have all of the information it needs to give the answers in such cases.

Once such an optimization is in place, then for C-like enums all of the cases would just fall through to the catch-all (i.e. there would be no preceding arms).

[steveklabnik]

Triage: no idea.

[steveklabnik]

Triage: same as in 2016

[Mark-Simulacrum]

I'm going to close this -- this essentially I believe comes down to this if -- which I think we can't remove in a completely backwards compatible manner today. If I'm wrong then we can reopen (but it becomes a really easy bug then I suspect, so it's odd that this wasn't fixed at the time).