Instead of finding the one pair with the highest frequency and merging it at each step, do the highest N pairs

I've been experimenting with my own little BPE implementations in other programming languages.

It seems that major bottlenecks are counting the frequencies of the pairs each iteration and merging the most frequent one.

I've noticed that instead getting the top two most frequent pairs and merging them each iteration is already a noticeable speedup for me with no noticeable loss of quality.

It should be possible to scale this up quite a bit before hitting diminishing returns or a drop in quality.