Context

Because of some type problem we need multiples times in the codebase to duplicate function definition in the Computional Backend part.

This PR fix this by doing a refactoring on the type part. The new one is less precise and correct that before but more flexible and easier to maintain.

Main difference are two folds:

• Use Instance instead of class in get_comp_backend (to fix the static method pb)
• The return type of the function in the computational backend is always np.ndarray or torch.Tensor. Therefore mypy cannot now that if you passed a TorchTensor (from us) the method will return a TorchTensor. This is a limitation but it allow us to remove the second TypeVar Generic in the Computational backend. This means in some case we need either to tell mypy we know the return type will be a TorchTensor by calling cast or TorchTensor._docarray_from_native()