Public tokenizer errors, hasChatTemplate

[pcuenca]

See ml-explore/mlx-swift-examples#181 (comment), ml-explore/mlx-swift-examples#150

In addition to making the tokenizer errors public, this PR adds a new hasChatTemplate property. We could also create an accessor for the chat template itself if needed, but we still need to support array values that are in the process of being deprecated anyway.

Would these changes suffice for mlx-swift, @davidkoski? Do you think we'd need something like Python's tokenize=False argument, like @awni mentioned?

[davidkoski]

Would tokenize=False be a method that just applied the template and returned the string? The only thing I can think we need it for today is if somebody wanted to log it for debug purposes -- perhaps the llmtool would call it and log it.

Right now the only problem I know of is the caller not knowing if they can safely call the applyChatTemplate method itself, so public error and hasChatTemplate seems sufficient.

[pcuenca]

Yes, I think there are interesting use cases for tokenize=False, I've seen people decoding the ids as a workaround.

I'll merge this then and we can incorporate tokenize=False later. Thank you!