This improves the runtime of `Enum.dedup_by/2` on lists by replacting the generic `reduce/3` implementation with a hand-rolled version similar in spirit to the `dedup_list/1` optimization. In my benchmarking, the magnitude of the improvement depends heavily on the size and shape of the data, but it looks roughly like this:
| List length | Duplicates | Speedup |
| 10 | 0 | +140% |
| 1,000 | 0 | +81% |
| 1,000,000 | 0 | +7% |
| 10 | 20% | +116% |
| 1,000 | 20% | +64% |
| 1,000,000 | 20% | +2% |
| 10 | 50% | +45% |
| 1,000 | 50% | +62% |
| 1,000,000 | 50% | +12% |
Tyler A. Young
05576dd133