Add the interface for fix/unfix parameters for gradient optimization.

[zhenwendai]

[marpulli]

I keep forgetting how to fix things so I might have a go at this. How would you do it? A method on the Inference object? Something like: infr.fix_parameter(m.X)?

[meissnereric]

Looks fine to me. With another call to InferenceParameters inside that, similar to fix_all but only fixing a single parameter.

One question I have is around the difference between model and posterior parameters here. Are they any different in this context? I'll give it some thought in the meantime, but anything you or @zhenwendai can think of?