Broadcasting issues during Backward

[nathanielsimard]

When doing broadcasting during the forward pass, sometimes it can cause error during the backward pass, especially when broadcasting feature dimensions with the batch dimension. We should add automated tests that support those usecases.

Example:

let w = Tensor::<B, 1>::zeros([16]).require_grad();
let x = Tensor::<B, 1>::zeros([4]).require_grad();

let y = x.slice([0..1]) * x;
y.backward() // panic!

We should add tests for all binary operations that support broadcasting.