Scalecast

Added
- Added `util.metrics.abias()` function.
Changed
- Revamped the RNN model. Now accepts exogenous regressors and other small changes that I believe will make it faster and more accurate, as well as allow for more customization.
Fixed
- Fixed the `Pipeline.backtest()` function, which was adding too much space between consecutive training sets.
- Fixed the train_only argument in `SeriesTransformer.DetrendTransform()` when `loess = True`.

Added
- Added a LOESS detrender (from statsmodels) to the `SeriesTransformer` object. Called using `SeriesTransformer.DetrendTransform()`.
Changed
- `SeriesTransformer` can seasonally adjust the same series multiple times (to capture multiple seasonalities).
- The `util.find_optimal_transformation()` function now tries a LOESS detrender by default.
Fixed

Added
Changed
- "Differenced" no longer displayed when calling `Forecaster.__repr__()`.
Fixed
- Fixed plotting confidence intervals on separate axes in the `Forecaster` and `MVForecaster` objects.

Added
- Added `Forecaster.STL()` method.
- Added `cilevel` argument to `Pipeline.backtest()` method.
- Added `SeriesTransformer.DeseasonTransform()` and corresponding revert function.
- Added `util.metrics.bias()` function.
- Added `seasonal_adj` in default `try_order` argument from `util.find_optimal_transformation()`.
Changed
Fixed
- Fixed how metrics are calculated after reverting a detrend transformation.

Added
- Added a `Forecaster.add_metric()` and `MVForecaster.add_metric()` argument that can be used for custom metrics.
- Added a default `called` argument in `MVForecaster.add_optimizer_func()` to be the function's name.
Changed
- Changed how metrics are set when initializing the objects and how they can be set later.
- Changed how metrics can be accepted in `util.backtest_metrics()`.
Fixed
- Fixed the error message that is displayed when setting metrics goes wrong.

Added
Changed
- Removed pandas-datareader as a dependency.
Fixed