Propagate variances correctly in dependency expressions

[henrikjacobsenfys]

General

As mentioned in #291, variances are not propagated properly in dependency expressions. "Just reduce your expression correctly" does not seem like a good solution, especially when it's quite common to write a*a rather than a**2.

Idea

When adding more math functions (#297 ), we can also include their derivatives, and thereby (quite straightforwardly and generally) add support for more proper variance propagation. A simple check of parameter names when resolving dependency expressions can handle expressions like a*a*a, or a + (1 - a) correctly.

This will be turned off during a fit for speed, but will be used in the final step of the fit when the parameters are updated, to propagate variances properly.