This was pretty obscure.  elemLocalRdrEnv was utterly wrong (replied
False when it should reply True) when given an Exact Name. That
doesn't happen often, but it does happen in the result of a TH splice.
The result was that an associated type didn't get a type variable that
lined up with its parent class (elemLocalRdrEnv is used in
RnTypes.bindHsTyVars), and that messed up the singletons package.

I've made a completely different test case to show up the bug:
indexed_types/should_fail/T9160

I also refactored RdrName.LocalRdrEnv to be a record with named
fields, which makes the code more robust and easy to understand.