fix CompositeImplicitAutograd ops improperly labeled (#69169)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/69169
I checked `derivatives.yaml`, and it doesn't look like `logical_not/and/xor` are meant to work with autograd. Those 3 ops are currently set as `CompositeImplicitAutograd` though, implying that they do work with autograd. Updating them to be CompositeExplicitAutograd instead.
This came up because I'm trying to improve the error checking in external backend codegen, and these ops being improperly labeled incorrectly triggers my new error checks for XLA (see https://github.com/pytorch/pytorch/pull/67090)
Test Plan: Imported from OSS
Reviewed By: zou3519
Differential Revision: D32739976
Pulled By: bdhirsh
fbshipit-source-id: a756dd9e0b87276368063c8f4934be59dca371d3