New abstractions for rolling-window forecasters.

[Emerald01]

This commit introduces new abstractions for sklearn-style models (i.e. they support fit() and predict()) which predict future values based on a fixed historical window.

1. RollingWindowDataset supports __iter__ and __getitem__ which process data to the required (past, future) format
2. SKLearnForecaster(ForecasterBase) handles (1) univariate: the hybrid (multioutput + autoregression) train and forecast (2) multivariate with prediction_stride = 1: autoregression train and forecast (3) multivariate with prediction_stride > 1: set prediction_stride = max_forecast_steps, and directly return the model's prediction of the target univariate for the next max_forecast_steps timestamps
3. minor: remove sample_mode, I feel the stats sampling mode is never used and complicates the interface
4. Unittests pass for all these modes mentioned, performance-wise and result-wise, same as before,
   for example, python test_boostingtrees yields the following logs:

2022-10-19 14:22:30,957 (resample:144) WARNING: Inferred granularity 0 days 01:00:00
2022-10-19 14:22:31,282 (resample:144) WARNING: Inferred granularity 0 days 01:00:00
2022-10-19 14:22:31,570 (resample:144) WARNING: Inferred granularity 0 days 01:00:00
--------------------------------------------------------------------------------
2022-10-19 14:22:31,736 (test_boostingtrees:83) INFO: test_forecast_multi_autoregression
--------------------------------------------------------------------------------
2022-10-19 14:22:31,853 (sklearn_base:124) INFO: Model is working on a multivariate dataset with prediction_stride = 1, model will be trained to autoregressively predict all univariates.
2022-10-19 14:22:36,791 (test_boostingtrees:89) INFO: Immediate forecast RMSE: 2.95
2022-10-19 14:22:36,836 (test_boostingtrees:97) INFO: Look-ahead RMSE with time_series_prev: 18.95
.--------------------------------------------------------------------------------
2022-10-19 14:22:36,921 (test_boostingtrees:109) INFO: test_forecast_multi_seq2seq
--------------------------------------------------------------------------------
2022-10-19 14:22:37,030 (sklearn_base:130) INFO: Model is working on a multivariate dataset with prediction_stride > 1. Model will directly forecast the target univariate for the next 2 timestamps.
2022-10-19 14:22:40,525 (test_boostingtrees:115) INFO: Immediate forecast RMSE: 3.61
2022-10-19 14:22:40,567 (test_boostingtrees:123) INFO: Look-ahead RMSE with time_series_prev: 19.55
.--------------------------------------------------------------------------------
2022-10-19 14:22:40,627 (test_boostingtrees:135) INFO: test_forecast_uni
--------------------------------------------------------------------------------
2022-10-19 14:22:40,675 (sklearn_base:118) INFO: Model is working on a univariate dataset, hybrid of sequence and autoregression training strategy will be adopted with prediction_stride = 2.
2022-10-19 14:22:43,828 (test_boostingtrees:142) INFO: Immediate forecast RMSE: 1.47
2022-10-19 14:22:43,854 (test_boostingtrees:150) INFO: Look-ahead RMSE with time_series_prev: 17.27

[salesforce-cla]

Thanks for the contribution! Unfortunately we can't verify the commit author(s): Tian Lan <t***@t***.i***.s***.com>. One possible solution is to add that email to your GitHub account. Alternatively you can change your commits to another email and force push the change. After getting your commits associated with your GitHub account, sign the Salesforce.com Contributor License Agreement and this Pull Request will be revalidated.

[aadyotb]

Thanks for the effort Tian! In addition to the changes mentioned inline, can you update the docs/source folder to ensure that the new files are included in the API doc?