improves huggingface_hub example (#988)
The provided example uses the default `max_length` of `20` tokens, which
leads to the example generation getting cut off. 20 tokens is way too
short to show CoT reasoning, so I boosted it to `64`.
Without knowing HF's API well, it can be hard to figure out just where
those `model_kwargs` come from, and `max_length` is a super critical
one.
