Potential unintended shape inconsistency with permute.

[ppaleja]

When working on #68, I came across this curiosity when edge testing different autograd calculations. Consider this code:

let a = cx.tensor((2, 1)).set([[1.0], [2.0]]);
let a: GraphTensor = a.expand((2, 2));

Since a is now a view with shape (2, 2), I should be able to permute with

let a = a.permute((1, 0));

or something similar.

But this code fails. I can only permute, by knowing the stride length (which I could only know if I was keeping track of shape tracker, i.e. seems like bad encapsulation)

let a: GraphTensor = a.permute((0, 2, 1));

Not only this, but

// 4. Sum to a scalar loss
let loss = a.sum((0, 1)); // Sum over logical axes 0 and 1

// 5. Autograd
let grads = cx.compile(Autograd::new(orig_a_id, loss), ());
cx.keep_tensors(&grads);
cx.execute();

Will cause an OOB error.

I couldn't find any equivalent pytorch methodology, so I don't know what the convention would be for this (if this behavior is intended), but it seems like a vulnerability to me.