Potential performance improvement: cache the result of mergeBytePairs()

[mikethea1]

When tokenizing large input strings, mergeBytePairs seems to be the bottleneck even when it isn't [degrading to quadratic[(https://github.com//issues/25).

On my workload, I found that a small change of caching the result of mergeBytePairs resulted in a ~33% speedup since larger files tend to have many repeat pieces.

This isn't a free improvement since it adds some overhead due to checking the cache.

Have you considered this optimization before? Would you entertain a PR to add it?

[yethee]

You can submit a PR and we can run a performance test against the current version.