[WIP] Use unboxed sums for change / no-change tracking

[sjakobi]

…instead of pervasive ptrEq.

[sjakobi]

I suspect that we're less lazy in the result of two now, and that this is the actual source of the minor performance improvements.

[sjakobi]

I should try to make two itself more strict: Force the result of the recursive application of go, force the returned BitmapIndexed nodes.

[sjakobi]

With the benchmarks I'm getting a lot of noise and a rather reliable 2-3% speedup in HashMap.insert-dup.ByteString.

[treeowl]

The last time I experimented with this idea, the performance was worse, but maybe things have changed. Or maybe there are differences between HashMap and Map that explain the difference.

[treeowl]

FWIW, I think those old experiments used unboxed tuples with Int# tags rather than the (then bleeding edge) unboxed sums, but I don't know why that would make a difference, considering the way GHC implements unboxed sums.