[testing] fixing crash in deberta

[stas00]

This PR fixes a crash in deberta tests w/ pytorch-1.7+:

RuntimeError: diff_view_meta->output_nr_ == 0 INTERNAL ASSERT FAILED at "/opt/conda/conda- \ 
bld/pytorch_1603436966316/work/torch/csrc/autograd/variable.cpp":363, please report a bug to PyTorch.

All credits go to @gchanan, thank you! For details of why please see #8022 (comment)

Fixes: #8022

[LysandreJik]

Great, thanks @stas00 and @gchanan!

[stas00]

p.s. apparently there has been a deprecation warning with pytorch-1.6, but since this project isn't quite paying attention to the warnings it lead to this crash with pytorch-1.7. It seems to be a weird situation where a deprecation warning hasn't been turned into an error and instead leading to a crash with this particular issue, but perhaps setting an intention to keep the warnings in check would save a possible hassle in the future. #8060

Perhaps what would help is to automatically turn selective types of warnings to errors, and thus not let those slide until they become a problem.

Also having a scheduled CI that runs occasionally on pytorch-nightly (and any release candidates) would give an early alert.