Bayesloop

Fixes:
- restore description on pypi

New features:
- New transition model: Bivariate random walk

Fixes:
- various import fixes
- more stability for complex transformations in the `Parser` module

Development:
- moved tests and coverage to Github Actions

New features:
- New observation model: Laplace distribution
- Hyper-parameter optimization now supports "forward-only" algorithm

Fixes:
- Model evidence of `ChangePoint` transition model depended on the chosen grid-size
- `RegimeSwitch` transition model did not support integer parameter values
- Jeffreys prior for Gaussian observation model was parametrized on variance, not standard deviation
- `SymPy` observation models now support Beta function

New features:
- Additional API functions in `OnlineStudy`
- Probability Parser for arithmetic operations on inferred (hyper-)parameters
- Custom likelihood functions (observation models) based on NumPy functions
- Universal `plot` method
- Convenience methods `load`, `set`, `add`, `eval`

Fixes:
- Support for `besseli` function in SymPy models
- Consistent order of parameters in SymPy/SciPy models
- Consistent order of parameters in joint-distribution plots
- Fix to support SymPy 1.1
- `AlphaStableRandomWalk` transition model
- `NotEqual` transition model

Development:
- `bayesloop` now features automatic testing based on `TravisCI`.
- Automatic code coverage evaluation by `coveralls.io`

Fixes
- Hotfix for scaling of hyper-prior values in `ChangepointStudy`, resulting in distorted model evidence values. This bug was introduced in version 1.2.0.

Fixes
- Use relative imports only, thereby adding support for Python 3.6