Scalecast

Added
- Added more feature importance options, all sourced through the shap library.
Changed
- shap is now a requirement and eli5 is not.
- Changed `Forecaster.reduce_Xvars()` to use only shap feature importance to rank features.
- Removed `fi_method` argument from `tune_test_forecast()`.
Fixed
- Fixed how a pandas function was called that was raising a warning.
- Fixed feature importance to use shap only with TreeExplainer, PermutationExplainer, and other explainers (85). See the [docs](https://scalecast.readthedocs.io/en/latest/Forecaster/Forecaster.html#src.scalecast.Forecaster.Forecaster.save_feature_importance) The eli5 package appears to be deprecated.

Added
- Added `'verbose':[-1]` to lightgbm grids.
Changed
- Refactored code and changed library version requirements to eliminate warnings that were originating from the seaborn (49) and pandas libraries.
Fixed

Added
- Added `set_aside_test_set` argument to the `util.find_optimal_transformation()` function to prevent leakage.
- Added the `util.gen_rnn_grid()` function.
- Tensorflow models now save the `ymin` and `ymax` attributes in history to facilitate transfer learning.
Changed
- Changed where `boxcox_tr()` and `bocxcox_re()` functions are located in `src` so that `Transformer` and `Reverter` objects can more easily be pickled and the output when printed is more descriptive.
- Changed the lightgbm default grid to suppress output.
Fixed
- Added `**kwargs` in `util.backtest_for_resid_matrix()` to make pipelines that require kwargs work.

Added
- Added transfer learning for RNN and LSTM models.
- Added more Xvars to infer in the `util.infer_apply_Xvar_selection()` function.
- Added `regr` argument to `Forecaster.transfer_predict()`.
- Added `Forecater.save_tf_model()` and `Forecaster.load_tf_model()` functions.
Changed
- Transferred models store Xvar info in history.
Fixed
- Fixed an issue with RNN models where AR terms that were not sequential were not counted correctly.
- Fixed an issue where Xvars were not being recorded correctly in history for RNN models.

Added
- Added `Forecaster.transfer_predict()` method. Only univariate sklearn models supported for now (77).
- Added `Forecaster.transfer_cis()` method.
- Added `carry_fit_models` attribute in `Forecaster` object that can be changed when object is initialized.
- Added `util.infer_apply_Xvar_selection()` function.
Changed
- Changed how many history attributes are stored for each evaluated model, making the `Forecaster` object more memory efficient.
- Refactored forecasting code for sklearn models so that model evaluation is more efficient.
- Changed the `max_ar = 'auto'` behavior in `Forecaster.auto_Xvar_select()`.
- Changed scikit-learn dependency to `<1.3.0` due to it not working with the shap library.
Fixed
- Fixed an issue with combo modeling where defaults were not working when a previous model had been run test only.

Added
Changed
Fixed
- Fixed an issue where `impute_lookback=None` was not working with `fill_strategy='moving_seasonal_average'` (70).