[forms] Implement gradient of TransformedOperator

Since a transformation is linear, applied pointwise, and independent of the position, we can in principle exchange the order of transformation and derivative. However, we do in general not now an appropriate type for the jacobian. E.g. for matrix valued functions we would need a third order tensor.

However, when viewing the gradient as the transposed of the jacobian, it is a vector of partial derivatives, where the outer most index corresponds to the partial derivative direction while each entry has the same type as the range of the original function. Hence we can generically use std::array<Range,dim> as range for the gradient of a function with range of type Range.

This e.g. allows to transform a power function space to matrix valued finite elements using a local isomorphism and then to compute gradients of this space. To make this usable, this also adds an overload of dot(x,y) for two std::array arguments. To make this more general, one may want to implement more overloads for different combinations of types or more generic ones later.