update the dispatch of some operators which accept scalar (#110918)
The scalar overloads of some ops like `bitwise_xor.Scalar` were dispatched to `CompositeImplicitAutograd` by default. It is against the rule for `CompositeImplicitAutograd` that all tensor operations (except reading metadata) must be done through calls to the ATen dispatcher rather than interacting with the Tensor directly.
So, here update the dispatch of these overloads to `CompositeExplicitAutograd`.
Fixes #93224
Pull Request resolved: https://github.com/pytorch/pytorch/pull/110918
Approved by: https://github.com/peterbell10
