Adding requires_grad flag to Value class

[rhalbersma]

I've added a requires_grad flag to the Value class that defaults to False, mimicking PyTorch's behavior.

The topological sort of the computational graph can then stop expanding nodes that don't require gradient propagation. E.g. in expressions such as x + 2 the conversion of the scalar 2 to Value(2) will no longer propagate the gradient for that node.

I've also incorporated these changes into the neural network where all parameters receive requires_grad=True. All unit tests continue to pass and the demo notebook finds the exact same loss after 99 iterations.

Note that I also did some other small refactorings such as the separation between local gradient computation and propagation. This made it easier to selectively propagate the gradient based on the requires_grad flags. I've renamed the internal _backward attribute to a constructor argument gradient_fn.

All in all, the engine.py module was reduced by 15 lines, and weighs in at less than 80 lines now.

[rhalbersma]

Any update on this? Is this repo not taking PRs, or should I polish the PR a bit more?

[fcakyon]

@rhalbersma unfortunately this repo doesn't take PRs..