Remove native_functions.yaml dependency from DistributionBernoulli.cu (#67721)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/67721
This moves the operator calls (`expand_inplace` and `to(kCUDA)`) into
`bernoulli_impl_` which is shared between CPU and CUDA. So that the
cuda code only needs to generates random numbers and nothing else.
The other changes are just rearranging includes to avoid including
`Tensor.h`.
Test Plan: Imported from OSS
Reviewed By: mruberry
Differential Revision: D32500426
Pulled By: malfet
fbshipit-source-id: f855c2055392355e79e3df832fa56d2041ebf24e
(cherry picked from commit 7ab7d17e941a6fb10be03f9a9538bc66f195b95f)