Add support for automatic differentiation

[rjpower]

Since we're capturing the expression graphs, we're in a good position to support automatic differentiation as well. This would require expressions to track a bit more information (the actual arithmetic operation being performed).

[rjpower]

@fegin, if you finish your other stuff, you can start looking into this.

It might require that we defer converting user expressions to Map/Reduce style operations immediately, and instead keep them in a higher form. I'm not sure if this is the case.

The basic idea is to use the chain rule/multiplication rule to take expression like x * w, and allow the user to ask the derivative with respect to w:

y = x * w
yprime = y.derivative(w)
# yprime should equal "x"

For a given operation, say Map, and assuming it takes 2 arguments, we compute the derivative by returning something like:

MapExpr:
  def derivative(self, dx):
    a, b = self.children
    return a.derivative(dx) * b + b.derivative(dx) * a