Make `ModelTokenizer` a `Transform`

[RdoubleA]

When you pass in a torchtune.modules.tokenizers.ModelTokenizer to torchtune.datasets.SFTDataset as the model_transform, pyre will complain because it is not typed as a Transform. We don't run into any runtime errors because all our tokenizers usually double inherit from ModelTokenizer and Transform (e.g., see Llama3 Tokenizer), but if I were to put together a custom dataset, the typing would be incorrect:

def python_code_instructions_alpaca(tokenizer: ModelTokenizer) -> PackedDataset:
    """
    Python code instruction-input-output pairs from iamtarun/python_code_instructions_18k_alpaca templated with Alpaca.
    """
    ds = SFTDataset(
        model_transform=tokenizer,
        source="iamtarun/python_code_instructions_18k_alpaca",
        message_transform=AlpacaToMessages(
            train_on_input=False,
        ),
        # pyre-ignore[6]: Incompatible parameter type
        split="train",
    )
    if tokenizer.max_seq_len is None:
        raise ValueError(
            "PackedDataset requires a max_seq_len to be set on the tokenizer."
        )
    return PackedDataset(ds, max_seq_len=tokenizer.max_seq_len, split_across_pack=False)

This is not an issue on our linters as we don't run pyre (I believe?), but this typing is technically incorrect and should be adjusted as some point.

[thomasjpfan]

I think the typing for SFTDataset.model_transform is correct and does not need to be adjusted. SFTDataset requires model_transform to have __call__ defined and not tokenize_messages.

If a user's ModelTokenizer object also follows the Transform protocol, they have two options:

1. Create a new protocol to represent the combination:

class ModelTokenizerTransform(ModelTokenizer, Transform, Protocol):
    pass

def python_code_instructions_alpaca(tokenizer: ModelTokenizerTransform) -> PackedDataset:
    ...

2. Use Transform if the function does not require tokenize_messages.

def python_code_instructions_alpaca(tokenizer: Transform) -> PackedDataset:
    ...