[JIT] fix opinfo utils to handle tensor kwargs
Previously, the traced function would only take tensor args (and not tensor kwargs). Then the tensor kwargs would be inlined as constants into the graph during tracing, and changes to the kwarg inputs wouldn't be reflected in new calls to the traced function.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/76768
Approved by: https://github.com/mruberry
