Implement torch_tensor_backward

[jwallwork23]

Closes #152.

This PR:

• Sets the requires_grad argument to tensor constructors up properly. (Done in Hook up requires_grad #288.)
• Wraps the tensor->backward method as torch_tensor_backward.
• Provides a method for retrieving the gradient of a tensor (with respect to a tensor on which the user has called torch_tensor_backward): torch_tensor_get_gradient.
  • For similarity with PyTorch, I also provided a method torch_tensor%grad() (c.f. tensor.grad).
• Provides unit tests for the gradients of the overloaded operators.
• Updates the autograd example accordingly.
• Revises the scalar multiplication/division approach as described below. (Done in Rework scalar multiplication and division #289.)

I spent far too much time trying to get the gradients to be computed correctly for scalar multiplication and division as it is currently implemented. However, I discovered that everything works nicely if we just define scalars as rank-1 tensors with a single value. The revised approach to scalar multiplication/division adds a bit of boilerplate, but perhaps this could be reduced by introducing a torch_scalar class (see #285 for discussion).

[jwallwork23]

I'll break this PR down into more manageable ones as it's got rather unwieldy.

[jwallwork23]

Ah, just noticed that I haven't updated pages/autograd.md for this PR yet

### The `backward` operator

*Not yet implemented.*

Will do that.

[jwallwork23]

Ah, just noticed that I haven't updated pages/autograd.md for this PR yet

### The `backward` operator

*Not yet implemented.*

Will do that.

Done in 883654a.