XLNet `use_cache` refactor (#5770)
Slightly breaking change, changes functionality for `use_cache` in XLNet: if use_cache is True and mem_len is 0 or None (which is the case in the base model config), the model behaves like GPT-2 and returns mems to be used as past in generation. At training time `use_cache` is overriden and always True.
