Change how cuda available memory is calculated in largeTensorTest decorator (#72207)
Summary:
Related PR https://github.com/pytorch/pytorch/issues/45332
Related discussion https://github.com/pytorch/pytorch/pull/45332#issuecomment-985996064
Pull Request resolved: https://github.com/pytorch/pytorch/pull/72207
Reviewed By: ngimel
Differential Revision: D34387921
Pulled By: mruberry
fbshipit-source-id: 2d842a25a5d3d1fc48917ba8fb29ff96d7bc2650
(cherry picked from commit 01a9e980c7745e4f8b7c983f403083deb7993f9a)
