Generating Dirichlet samples using the method based on samples from
the gamma distribution can result in samples being nan if all the
values in alpha are sufficiently small.  The fix is to instead use
the method based on the marginal distributions being the beta
distribution (i.e. the "stick breaking" method) when all values in
alpha are small.

* Create a struct for each random variate generation method:
  DirichletFromGamma and DirichletFromBeta.
* Move the initialization of the data required by the generators
  into the new() method.
* Make the main Dirichlet object an enum.

The main change is that Dirichlet now has a const trait N for
the dimension.  The other significant change is to propagate
errors that could occur when Beta:new or Gamma::new is called.