Unnecessary calls to .contiguous() hurt performance

[jacobhinkle]

Describe the bug
There are many times in our python code where we force tensors to be contiguous. This is because before pytorch introduced packed_accessor it was pretty annoying to use strided tensors when writing extensions

Expected behavior
We should be able to run most operations on strided tensors without copies.

Solution
We just need to convert most of our kernels and calling code to use packed_accessor32.