Align batch logits processor token contract

[neilmehta24]

After #1072, the data sent to the batched logits processors changed:

1. The type of self.tokens[e] is now a python list[int], not mx.array. This means that the logits processors receive python lists.
2. The current input token is now appending to self.tokens after the processor call, not before.

I believe this is a bug for two reasons:

• For the non-batched API, processors still get an mx.array, and the current input token is appended before the processor runs.
• There's a comment that says # Update the token context that will be used by the logits processors, but see how _token_context is not being used within the logits processors block.

This PR brings the data sent to the logits processors in line with the pre-refactor data.

[angeloskath]

Thanks for catching that. That's a typo actually. They should be receiving self._token_context[e] that's just above. This also limits the max context to 256 tokens which is fairly arbitrary and different than the non-batched API. Let me know if you think that's a problem. As you see it is fairly easy to change.