Rethink the core overlap computation for the tokenizers -- AIList and speed tokenizers

[nleroy917]

The current tokenizers use rust-lapper as an overlap computation method. This is nice but we should be using our own algorithms -- namely AIList. Moreover, tokenization can become even faster if we make some assumptions about our data, like is it sorted?

Things we should decide on:

1. Can we replace rust-lapper with AIList?
2. Can we create a version of the tokenizers that assumes sorted files? Call it a SpeedTokenizer
3. Should the above SpeedTokenizer be checking for sorted-ness?