Add aten::lift_fresh to the exception list that converts input to torch tensor

[barney-s]

This was causing fastNLP_BERT model to fail.

https://github.com/fastnlp/fastNLP/blob/4e95989e973f59b2ecb7f718647257e8b6fea0c7/fastNLP/embeddings/bert_embedding.py#L427

467                word_pieces[i, 1:word_pieces_lengths[i] + 1] = torch.LongTensor(word_pieces_i)

torch.LongTensor was returning a torch.Tensor instead of XLA Array.

Debugging the call trace

DISPATCH: aten::lift_fresh
 FUNCTION: aten::lift_fresh

fixes: #8126