I've made the changes to pivot selection discussed earlier in this PR, and double checked that this is not a breaking change (note that we only changed devDependencies in bower.json). Also, now that integers exports an int pow function, I've updated the tests to use that instead of defining it here using the FFI. I think this is good to merge now.

Did we benchmark it at all in the end?

We did, I added a comment in #48 with results. Performance wise it seems to be pretty much the same as the previous implementation, at least on V8.

@hdgarrood I'm glad it worked; it didn't in my own tests.

The asymptotics have several dimensions, each induced by particular applicatives. Some examples:

1. A Writer with O(n) append will multiply out to O(n²) with the original implementation, but be O(n log₂n) with divide-and-conquer.
2. A Writer whose monoid is fast when folding one way but slow when folding the other way will be somewhere in between. For example, a cons list monoid with the elements more-or-less distributed evenly will be O(n) with a right-folding traversal, O(n²) with a left-folding traversal, and O(n log₂n) with divide-and-conquer. A monoid that's slow when right-folding, but fast when left-folding, also ends up with divide-and-conquer as a middle ground.
3. For Reader there's no difference when traversing, and no performance difference when invoking the resulting function. But invoking the resulting function will consume O(n) stack for the original implementation (even if "stack-safe", hence the error added by the test), but O(log₂n) stack for divide-and-conquer. Note that to observe this, you have to invoke the function; the performance of calling traverse won't include this information.

Resolved by #78 (!)