MAINT: tweaks to prior constraints

[mj-will]

Some tweaks to prior contraints

• Raise an error if a constraint is present but the parameter key is missing. Currently, it's possible to have constraints that are silently ignored, e.g. because the conversion function doesn't generate them.
• Remove an unreachable if statement. Prior to Change probability initialization to use np.ones_like #1028, evaluate_constraints returned 1.0 if there were no valid constraints and there was a check for this where the normalization was set to 1. I've removed this for now. Alternatively we could return 1.0 if all values are true but the while should do this anyway.