Clean-up non-AVX variant of `bitwise_binary_op` template (#36966)
Summary:
Compute number of element as `constexpr` and use it as both `buffer` element size as well as for upper boundary
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36966
Differential Revision: D21150602
Pulled By: malfet
fbshipit-source-id: 581634565c54c7295f3b77c8dc86659d5cc4ce19