[timeseries] Update the RecursiveTabular model

[shchur]

Description of changes:
This PR contains various enhancements and bugfixes for the RecursiveTabular model.

Enhancements:

• Apply standardization to each time series (after differencing) using the standardize argument.
• Limit the # of rows in the tabular dataset for faster training using the max_num_samples argument.
• Obtain quantiles for the probabilistic forecast assuming zero-mean Gaussian distribution of residuals (multiplied by the scale of the respective time series). This ensures that the mean/0.5 quantile forecasts exactly correspond to the point forecast by the TabularPredictor.
• Allow passing hyperparameters and fit_kwargs to the TabularPredictor.
• Replace AutoGluonTabular with RecursiveTabular in the presets. The updated version achieves 76% win rate against the previous one in terms of MASE.
• (unrelated to the model) Speed up the implementation of get_forecast_horizon_index_ts_dataframe using vectorization.

Bugfixes:

• Make sure that the time series are long enough for the chosen differencing. Before, if some time series were too short, numba would crash the entire Python process.
• Fix a bug where MLForecast would return incorrect predictions if the item_ids were not sorted alphabetically.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[github-actions]

Job PR-3184-9a0caaf is done.
Docs are uploaded to http://autogluon-staging.s3-website-us-west-2.amazonaws.com/PR-3184/9a0caaf/index.html

[github-actions]

Job PR-3184-9ab0868 is done.
Docs are uploaded to http://autogluon-staging.s3-website-us-west-2.amazonaws.com/PR-3184/9ab0868/index.html