ENH improve HDPModel L-BFGS objective parameterization to avoid overflow

[Mike Hughes]

Currently, HDPModel needs to do gradient optimization of its factor

q( v_k | u_k0, u_k1) = Beta( v_k | u_k0, u_k1 )

we could instead parameterize the beta in terms of a "scale/sum" and a "mean". This would perhaps make gradient descent more stable.

[Mike Hughes]

Completed with commit of HDPModel2.py, which uses OptimizerForHDP2.py to solve a re-parameterized objective. Generally more stable in practice.