Rolch

This release includes
- a nice overhaul of the distributions backend by BerriJ
- sanity checks for distribution and link function compatibility
- sanity checks for the distribution support
- some new docs for building new links and distributions
- more functionality for the `debug` mode by simon-hirsch

Improving the scaler:
- Allow for scaling of certain variables only. See https://simon-hirsch.github.io/rolch/scaler/ for details. You can now pass either a bool into `scale_inputs` or a numpy array with indices of the columns of $X$ you'd like to scale.

Some fixes:
- Add MLE estimates as start values to improve convergence in the beginning. This is more robust, since the previous start values might have been overly optimistic.
- Allow for boolean arrays in the `equation` and the `scale_inputs`.

This release introduces the distinction between the `Estimator()` classes, which provides the `.fit()`, `.update()`, `.predict()` methods and the `EstimationMethod()` classes, which do the actual fitting of $X$, $y$ to coefficients / weights / betas. This gives

- easier integration of new estimation methods
- easier handling of default parameters for methods
- easier handling of non-default parameters for methods, _especially for non-standard parameters like bounds, regularization strengths, etc.._
- much cleaner code in `OnlineGamlss()`

Breaking change: This change gets rid of the `estimation_kwargs` parameter in `OnlineGamlss()` which was cumbersome and poorly documented anyways.

Furthermore, we
- align the API of `OnlineGamlss()` and `OnlineLinearModel()` and derive the `OnlineLasso()` from `OnlineLinearModel()` to show the flexibility of the new approach.
- Introduce a verbosity parameter for `OnlineGamlss()` to print information to the user.
- Add some properties to (slowly) align more to the `sklearn` API
- Some minor fixes like fixing 36
- Add a lot of documentation

- Add `OnlineLasso()` Estimator to docs and ensure basic functionality
- Fix bugs in `OnlineLasso()`
- Add some link derivatives and second derivatives of link functions

This release implements an important fix and does some maintenance and improvements in the backend.

Fixes
- Don't allow inversion of rank-deficit Gramian matrices - thanks to katche1010 for spotting.

Improvements
- Allow batch updates of Gramian and Inverse Gramians
- Proper naming of links and their derivatives

This release implements a more structured API for model estimation. We introduce the `equation` dictionary, which specifies the model for each distribution parameter. Additionally, we fit intercepts for each distribution parameter per defaul now.

Fixes issues 23, 22 by BerriJ, simon-hirsch.

Much appreciated feedback on the API design by Franz Kiraly.