torchdeploy allow embedded cuda interp use without cuda (#59459)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/59459
For any binary that can be used both with and without cuda, it's better to allow just including the cuda flavor of the interpreter. The previous logic would fail in this case, as it only allows using the cuda flavor if torch::cuda::is_available() reports true. Now, we unconditionally allow the cuda flavor to be used if it's present.
Test Plan: Added new unit test to exercise this scenario, ran locally on devvm without cuda.
Reviewed By: dzhulgakov
Differential Revision: D28902176
fbshipit-source-id: 5c7c90d84987848471bb6dd5318db15314e0b442
