Feature request: Fixing nb.theta during optimization

[willtownes]

As a user, I can request that glmpca not try to learn the negative binomial dispersion parameter during optimization by setting a flag in the control parameters. This issue is copied over from the python version: willtownes/glmpca-py#11