`RnTypes` contains a fairly intricate algorithm to extract
the kind and type variables of an HsType. This algorithm carefully
maintains the separation between type variables and kind variables
so that the difference between `-XPolyKinds` and `-XTypeInType` can
be respected.

But after doing all this, `rmDupsInRdrTyVars` stupidly just
concatenated the lists of type and kind variables at the end. If a
variable were used as both a type and a kind, the algorithm would
produce *both*! This led to all kinds of problems, including #13988.

This is mostly Richard Eisenberg's patch. The only original
contribution I made was adapting call sites of `rnImplicitBndrs` to
work with the new definition of `rmDupsInRdrTyVars`. That is,
`rnImplicitBndrs` checks for variables that are illegally used in
both type and kind positions without using `-XTypeInType`, but in
order to check this, one cannot have filtered duplicate variables out
before passing them to `rnImplicitBndrs`. To accommodate for this, I
needed to concoct variations on the existing `extract-` functions in
`RnTypes` which do not remove duplicates, and use those near
`rnImplicitBndrs` call sites.

test case: ghci/scripts/T13988

Test Plan: make test TEST=T13988

Reviewers: goldfire, simonpj, austin, bgamari

Reviewed By: goldfire, simonpj

Subscribers: rwbarton, thomie

GHC Trac Issues: #13988

Differential Revision: https://phabricator.haskell.org/D3902