Use consistent shapes of the observed proportions

[pawel-czyz]

Currently, we keep the observed proportions either as:

1. An array with shape (n_timepoints, n_variants) or
2. An array with shape (n_variants, n_timepoints).

The problem is that:

1. Different submodules use different convention (e.g., quasimultinomial and plotting submodules).
2. In fact, quasimultinomial submodule uses both conventions simultaneously!

As this is very counterintuitive and can lead to annoying bugs down the line (e.g., when calculating overdispersion), I'm annotating it as bug, rather than refactor.

[dr-david]

Indeed. Should we do n_timepoints t_timepoints ?