Bayesloop

Algorithm & API changes
- **`dill` module is a required dependency.** Loading and saving `Study` instances is no longer an optional feature.
- **Major refinements to `OnlineStudy`.** This type of analysis now behaves more like a `HyperStudy` and continually updates hyper-parameter distributions and transition model probabilities.
- **SymPy priors are not re-normalized.** This allows to define priors with a support interval that deviates from the defined parameter grid.
- **`get...Distribution()`-methods return probability values of (hyper-)parameters, not density values.** This allows easier post-processing of (hyper-)parameter distributions.

Fixes
- Hotfix for numerically stable computation of average posterior distribution (hyper-study)

Fixes
- fixed "dtype=object" error for likelihood array
- hyper-parameter values were not restored after fitting (hyper-study)
- improved verbosity settings (hyper-study and data import)
- evaluation of average posterior distribution now numerically stable for large data sets (hyper-study)
- hyper-study handles standard fits and simple change-point-analyses

Fixes
- import of list data
- use of custom timestamps in HyperStudy
- lattice scaling for deterministic transition models
- use of deterministic transition models with more than one hyper-parameter
- prevent excessive output in the case of inapt parameter boundaries
- progress bar also shown for standard fit method
- require updated version of `tqdm`, fixes warning

Fixes
- absolute model evidence value should not depend on parameter grid size; corrected scaling

New features
- Change-points can be used in serial transition models
- Added Gaussian mean model
- New API functions for online study class
- Added documentation for online study class