Address memory leak in overloaded operators

[jwallwork23]

Closes #295.

This PR makes torch_tensor_delete the destructor for the torch_tensor class. This means that it gets called automatically when a tensor goes out of scope. I've set it up such that the user may still call the subroutine explicitly if they wish, but this is no longer required.

I also changed the overloaded operators so that inputs and outputs to the C++ functions are all pointers, meaning that only the empty/zeros/ones/blob/array constructors create new torch::Tensor objects on the C++ side. The overloaded operators now merely operate on values for the data associated with the tensors they point to.

[jwallwork23]

Since the current best practice still works fine, I think we can defer mentioning the destructor here and put some text on that in the upcoming tensor docs in #291.

[jatkinson1000]

Yep, we discussed at length together yesterday and located the nullification together so I am happy with the concepts and approach.

A couple of non-blocking queries so happy for this to merge when ready.

[jatkinson1000]

Consider renaming test_tensor_constructors_destructors.pf?
Also I get why you test empty and delete in a single subroutine but would it be relatively straightforward to separate them to try and keep things monotised (is that a word...?)?

[jwallwork23]

Done in 0ac7b35 and b6042cb.