Moving Remove Tensor Type Specializations to after custom passes
This is to allow for Intel folks to use type information in their custom passes.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/71748
Approved by: https://github.com/eellison
