Support TikToken

[abhishekbhakat]

Kimi K2 Instruct uses a tiktoken tokenizer.
There are many models that supports that. It would be a nice to have model.

[olegklimov]

Current models are practically context-infinite, so the importance of tokenizers are not as high.

Back in a day we had to pack everything to 16K tokens so it was important. Now you can divide chars/3 and here is your token estimation 😁