Benchmarks don't seem to exercise collision handling

[sjakobi]

We currently have benchmarks with keys of type Int, ByteString and String. All of these use the instances provided in Hashable which I'd expect not to produce a lot of hash collisions.

For tuning the code that handles collisions, e.g. #435, it would be useful to have a key type that is more collision-prone.

[treeowl]

Historically, the view has been that collisions had better be rare enough not to matter much. There's plenty of O(n^2) gunk in collision buckets.

[sjakobi]

Yeah, this isn't a high priority task. It's mostly the result of me wanting to try the searchSwap trick in #435. Now I simply don't have any benchmarks to tell whether the new code is actually better or not. ;)