Allow variables in log_likelihood group to have different dimensions

[sethaxen]

According to the InferenceData spec, the log_likelihood group is unique in that the dims or coords of a variable can be different from its dims or coords in observed_data, prior_predictive, and posterior_predictive. e.g. if the likelihood of a variable y is a multivariate normal, then in observed_data, the shape might be (nchains, ndraws, length_y), while in log_likelihood, the shape would be (nchains, ndraws). Currently from_namedtuple and from_mcmcchains don't handle this correctly.