Handle missing chat template and propagate other errors

[DePasqualeOrg]

The silent error made debugging more difficult.

[davidkoski]

This catch was added for the case where the tokenizer didn't have a chat template at all and to give it something to fall back on. At the time TokenizerError wasn't public so we couldn't catch it, but now it is and we want this catch to only catch this case:

        guard let selectedChatTemplate else {
            throw TokenizerError.chatTemplate("This tokenizer does not have a chat template, and no template was passed.")
        }

What do you think about doing this?

        } catch TokenizerError.chatTemplate(let message) {
            if message == "This tokenizer does not have a chat template, and no template was passed." {
                // fallback case
                let prompt =
                    messages
                    .compactMap { $0["content"] as? String }
                    .joined(separator: ". ")
                let promptTokens = tokenizer.encode(text: prompt)
                return LMInput(tokens: MLXArray(promptTokens))
            } else {
                throw TokenizerError.chatTemplate(message)
            }
        }

That way real errors get propagated as error and this fixes #150

[DePasqualeOrg]

That makes sense, although it would probably be better to have a dedicated error case for this type of error in swift-transformers and check for that case rather than the exact error message. I'll open a PR for that.

[davidkoski]

I agree, a dedicated error case would be ideal.

[DePasqualeOrg]

This is ready for review.

[davidkoski]

Looks great -- thank you for chasing down the dependencies!