Improve tokenizer testing

[pcuenca]

• Retrieve tokenized results from a dataset instead of adding to the repo.
• Maybe create script to regenerate the reference baseline.
• Add more cases.
• Verify edge cases.