A perfect Lambert reflector actually reflects light in a cosine distribution - that is, the amount of light per unit area reflected in any given direction $R$ is proportional to $N \cdot R$. The reason the radiance appears constant for all angles is that as the view direction moves away from the normal, the reflected light per unit area decreases, but the surface area per projected beam area (and hence per pixel) increases by the same amount - the cosine factors cancel out.
Given this distribution, to be energy conserving, the probability density for the reflected light in all possible directions cannot sum to more than 1. If you integrate the cosine function over the whole hemisphere around the normal, you get $\pi$, so this is the normalization factor you need.