Avoid boxing arrays

[treeowl]

Previously, we used a lot of things that looked like

runST (act >>= unsafeFreeze)

The trouble is that GHC can't unbox past the runRW# primitive
that runST is based on. So this actually allocates an Array
constructor which we will then throw away immediately.

The way to avoid this is to call runRW# manually to produce
a raw SmallArray#, then apply the Array constructor on the
outside.

[treeowl]

@Fuuzetsu, @m-renaud, I don't trust my ability to benchmark. Do you think one of you could try this out? My own benchmarks suggest that it's generally good, sometimes neutral, and (for reasons that are entirely mysterious to me) fairly bad for the fromList/long/String benchmark.

[m-renaud]

I'll be out of town until Tuesday next week, but I can take a look when I get back.

[treeowl]

We need to see if this is still necessary on GHC HEAD. @bgamari just recently implemented and merged a GHC patch to try to fix this sort of thing.

[sjakobi]

We need to see if this is still necessary on GHC HEAD. @bgamari just recently implemented and merged a GHC patch to try to fix this sort of thing.

I'll mark this PR as a draft then to signal that it's not currently waiting for a review.

[sjakobi]

I have rebased this PR on my sjakobi/unbox-arrays branch.

I'll take a look at the perf implications once GHC 9.2.2 is released.

We need to see if this is still necessary on GHC HEAD. @bgamari just recently implemented and merged a GHC patch to try to fix this sort of thing.

@treeowl In which GHC version was this work released?

[treeowl]

I don't remember. And the new stuff sometimes just defeats what we want to do. So everything needs to be pored over.

[sjakobi]

I have opened #375 to finish this.