Change in calculation of ELBO in line 146

[ashkush]

Change in calculation of j -= (q.pi - 1).dot(log_lambda) + logC(q.pi) , instead will be j -= (q.pi - 1).dot(log_pi) + logC(q.pi), as per the calculation given in book Machine learning and pattern recognition by Bishop

[cangermueller]

Thanks for spotting this! Can you do a before/after comparison? It it performs as well or better, feel free to send me a pull request.