Allow ops without tensor args if only fallback kernel exists (#19284)

Allow ops without tensor args if only fallback kernel exists (#19284)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19284

Instantiating a dispatch table previously only worked when the op had a tensor argument we could dispatch on.
However, the legacy API for custom operators didn't have dispatch and also worked for operators without tensor arguments, so we need to continue supporting that.
It probably generally makes sense to support this as long as there's only a fallback kernel and no dispatched kernel registered.
This diff adds that functionality.

Reviewed By: dzhulgakov

Differential Revision: D14931926

fbshipit-source-id: 38fadcba07e5577a7329466313c89842d50424f9