`torch::deploy` Autograd unit test failing when using a gpu enabled machine + using python 3.7/3.8

[PaliC]

This issue can be reproduced using docker with #233, otherwise, one a) needs to be running the unit test on a cuda enabled machine + on python 3.7/3.8. This error appears to be mitigated after upgrading to python 3.9, however, it is unclear why.