Restructure log-likelihood import

[gibsramen]

Currently, BIRDMAn assumes that the log-likelihood values will be the same dimensions of the table. For the multinomial model, we are fitting each sample instead of each sample-feature, so there are only N log-likelihood values (x chain x draws). Need to to address in importing.

Best way is probably to modify *_fit_to_inference and add in model attributes so we can do this easier in the default models (related to #8 and #9).