This method takes a given normal, and computes a random ray in the
unit-hemisphere described by that normal.
We use cosine-weighted importance sampling because it leads to better
convergence and is a nice micro-optimisation (from four trigonometric
operations to only two).