Fix ids and lidx

[rgiordan]

The fsm_util code uses lidx indices, even though it is used both by DeterministicVI and the LogProbability modules. This design relies on the ids parameterization and the lidx parameterization having the same indices as one another for all parameters that are touched by fsm_util.jl. Besides being brittle, we don't document what these different parameterizations mean nor that they need to overlap in this way. We also don't assert it anywhere.

A better design would be to pass a ParamType as an argument to the functions, and to refer to the ids appropriate to that type.