Inference problem

[TTitscher]

• introduce InferenceProblem as main way to interact with inference algorithms
  • removed MultiModelError which was used as such an interface, but was misleading in cases where there was only one model error
  • introduce special VariationalBayesProblem that only accepts normal distributions for the parameter prior and gamma distributions for the noise precision
• introduce NoiseModel that extracts terms of the model error and either builds a vector for VB or a log-likelihood term
• adapt variational_bayes to work with a list of vectors where the length of this list corresponds to the number of noise terms
  • removes all this split stuff
• add class LatentParameter with length N and start_idx and multiple LatentParameter are collected in the LatentParameters dict, where the key is a latent name (instead of access via index)
• provide a sampling example using pymc3 in "tests/sampling_example.py" to demonstrate the steps from an existing VariationalBayesProblem to sampling
• (Edit:) change bayes.vb.Gamma to only contain a single gamma distribution. variational_bayes now needs a list of bayes.vb.Gamma for multiple noise groups

[TTitscher]

As mentioned in one comment, I changed the latent definition such that we now need to provide the parameter_list directly instead of this (maybe) confusing key. That now looks like this

@joergfunger Do you want to have a look at it again? I have no hurry in merging this.