Caching of partial calculations

[spflueger]

The current Tensorflow code runs relatively slow (about 3x slower than pycompwa), which is related to the fact that tensorwaves is not caching any computations. Even the ones that that stay constant throughout the whole fitting/optimization phase (WignerD functions).

Some flexible and easy to use scheme is necessary to implement local caching.

Update: A lot of things have changed since. Now with the sympy + lambdify, the whole model can be optimized much easier. A separate method was added to the Model interface performance_optimize.