Evaluation of the gradient of a function at a point

Adding a function named Function::eval_derivatives or Function::eval_grad, working similarly to Function::eval, which computes the gradient of a function at a point by simply calling on ufc to evaluate basis function derivatives instead of basis function values.

The difference in calling different ufc functions is trivial in itself, but there is a lot of code managing parallell support which makes the dolfin eval code less than transparent, which is why I haven't just done this myself.

With this feature in place, it can be made possible to evaluate arbitrary ufl expressions point-wise without making a projection in python.