An `unordered_set` doesn't contain duplicates of the key_type, but it may contain "duplicates" in the less-granular equivalence relationship imposed by the transparent hasher/comparator on non-key_type objects. The new tests are copied from the existing tests for `set` and `map`. Thanks to Marshall Clow for highlighting this issue/use-case.
N.B.: I just copied the existing count_transparent.pass.cpp from set... but that gives us both Marshall's count_transparent.pass.cpp and Ruslan's count.transparent.pass.cpp. This is probably a terrible idea, so I'm open to suggestions for new test names.
It might also be a good idea for us to have a "neg test" proving that m.insert({Tuple2{1,2}, 3}) does not use heterogeneous lookup — it deliberately converts Tuple2 into Tuple3 and then looks up the Tuple3 according to the more granular equivalence relation.