Drop torch_tensor_to_array

[jwallwork23]

Closes #293.

Note #272 will need to be merged first as this branch builds upon it.

As discussed in today's meeting, torch_tensor_to_array is no longer needed. It was added because the overloaded mathematical operators created new torch:Tensors in C++, rather than working with pointers to Fortran arrays. Since that was addressed in #297, the expected (correct) result from torch_tensor_to_array can be achieved by creating the tensor from which you extract data from using torch_tensor_from_array applied to the array you want to extract into.

This is much neater and allows us to drop all cleanup (the remaining nullify statements) from the unit tests.

[jatkinson1000]

Thanks @jwallwork23 The code all looks sensible to me, but I have a couple of things I would appreciate if you can explain.

To summarise my understanding: instead of generating a tensor (pointer) that needs to then be mapped to an array, we instead require a user to 'pre-bind' an array to the output tensor of an arithmetic expression that will then be populated with any data resulting from computation. Is that correct?

It occurs to me here that all of our tests are for "correct" outputs, we do not test for errors. As discussed last week, this is not doable in pFUnit, but out of interest do you have an idea of what the outcome is if the user does not "pre-declare" the array associated with the output tensor?

Also, just to check my understanding, we do not call torch_tensor_delete in the autograd example at all now because it is cleaned up by the final method when going out of scope at the end of the program?

[jwallwork23]

To summarise my understanding: instead of generating a tensor (pointer) that needs to then be mapped to an array, we instead require a user to 'pre-bind' an array to the output tensor of an arithmetic expression that will then be populated with any data resulting from computation. Is that correct?

That's correct.

It occurs to me here that all of our tests are for "correct" outputs, we do not test for errors. As discussed last week, this is not doable in pFUnit, but out of interest do you have an idea of what the outcome is if the user does not "pre-declare" the array associated with the output tensor?

Well you can't use torch_tensor_from_array without having declared the array, no?

Also, just to check my understanding, we do not call torch_tensor_delete in the autograd example at all now because it is cleaned up by the final method when going out of scope at the end of the program?

Yes, I dropped it for simplicity since we can do that now.

[jatkinson1000]

Thanks. Already approved, but happy to re-approve after changes and feel free to merge as you please.