Wotan

FIXED:
- Bug in ``cosine`` which caused rounding errors in ``numpy.linalg.lstsq``

CHANGED
- API of ``transit_mask`` to use the keyword ``time`` (instead of ``t``)
- API of ``flatten`` to use the keyword ``mask`` instead of ``weights``, and avoid the requirement of inverting masks (see updated example, it is much simpler now)

ADDED
- ``lowess`` detrending now available in combination with ``transit_mask``

Added ``transit_mask`` to mask known transits during detrending

Added own lowess implementation instead of statsmodels dependency
- faster (based on numba)
- cleaner code (38 sloc instead of 678 sloc in statsmodels)
- convergence to defined level (default 1e-6) instead of a fixed number of robustifying rounds

Additional methods:

Fitting a model that is a sum of Gaussian bases:
- ``ridge`` Ridge regression (L2 loss, [Tikhonov regularization](https://en.wikipedia.org/wiki/Tikhonov_regularization))
- ``lasso`` LASSO regression
(L1 loss, [Least Absolute Shrinkage Selector Operator](https://en.wikipedia.org/wiki/Lasso_(statistics)), [Tibshirani (1996)](https://www.jstor.org/stable/2346178?seq=1#page_scan_tab_contents))
- ``elasticnet`` [Linear regression model](https://en.wikipedia.org/wiki/Elastic_net_regularization) trained with 50% L1 and 50% L2 norm regularization

- Bugfix: Time format is now ``float64``. Required to correctly process long time formats such as HJD=2458239.57352 (thanks to Paul Leyland)

- Faster speed for CoFiAM (thanks to Alex Teachey)
- Added new method `cosine` (thanks to Aviv Ofir)