A rebase of !6340 (closed), which was itself a rebase of !5068 (closed). I attached NOINLINE to dec and enc, hoping that doing so wouldn't affect inlining decisions and thus regress ghc/alloc perf the same way !6340 (closed) did.
It does not make sense to inline dec and enc; their implementation cannot be simplified further. Except for the obvious unboxing worker/wrapper split that we do for NOINLINE too, thankfully. Both dec and enc are very cheap to compute, despite their apparent code size. It should be an adequate guard against degenerate WordMaps as described in #20222 (closed)/#19820/#19414 (closed).