Implement gradient for UnbinnedNLL

[spflueger]

Implement the analytic gradient.

• Clarify how to use the gradient in the Estimator with Intensity which hides the backend (I think this comes back to the "ComPWA math language"). So currently I think it is necessary to have a graph structure that describes the computation, which can be converted to an actual computation. Same concept that tensorflow follows. This would solve many problems at once.
• Actually implement the gradient

Some tensorflow specific code might look like this (copied from amplitf):

for i, p in enumerate(float_pars):
     p.update(par[i])
with tf.GradientTape() as gradient:
    gradient.watch(float_pars)
    nll_val = nll(*args)
    g = gradient.gradient(
        nll_val, float_pars, unconnected_gradients=tf.UnconnectedGradients.ZERO)
    g_val = [i.numpy() for i in g]
    return g_val