Elo-grad

:sparkles: Features

- Compatibility with any (fully) Narwhals compatible dataframe library.
- `scikit-learn` is an optional dependency meaning the default installation is much lighter weight.
- `scikit-learn` compatible rating systems are now available via the `sklearn` submodule.

:boom: Breaking Changes

- Unix timestamp of games are passed as a column (specified by the date_col parameter) rather than the index, and are expected to be in seconds rather than nanoseconds.
- `scikit-learn` is not installed as a default dependency.

:sparkles: Features

* Support for L1 and L2 regularisation.
* Custom labels for plots.

:book: Documentation

* Documentation for regularisation.
* Google Analytics support for docs.

:sparkles: Features

- `PoissonEloEstimator` - a `pandas` and `scikit-learn` compatible Elo rating system for count data based on Poisson regression.

:book: Documentation

- Documentation for the Poisson Elo rating system.
- Corrections to the derivation of the connection between logistic regression and the Elo rating system.
- Add missing license reference for NBA data.

:book: Documentation

- Small improvements/better navigation from main page/README

:boom: Breaking Changes

- Remove `Optimizer.update_model`
- Return generator from `Optimizer.calculate_update_step`
- Add additional regressor arguments to `Optimizer.calculate_update_step`

:lady_beetle: Fixes

- Reset `k_factor` in `EloEstimator` when performing `GridSearchCV`

:sparkles: Features

- Support for additional regressors

:book: Documentation

- Documentation for additional regressors

:zap: Performance

- Add `lru_cache` for `Model.calculate_expected_score`