[Feature]: make token count accessible without loading the model

[joein]

What feature would you like to request?

The new token_count method currently requires a model to be loaded before using it.
However, the better approach might be to allow loading just a tokenizer without loading a model

Is there any additional information you would like to provide?

No response