ContrastsMatrix will silently accept non-unique levels

[kleinschmidt]

This came up on discourse: https://discourse.julialang.org/t/generating-dummy-variables-from-a-vector-of-strings/65507

We should enforce that the levels passed to ContrastsMatrix are all unique, because otherwise you're going to silently get non-sensical results (not least because the inverse-index mapping unique levels to numeric indices will be wrong).

I don't expect this to be a common situation but it prevents users from making mistakes while mucking around in the internals if they so desire.