Fix minor bugs in time series forecasting

[shchur]

Fix several minor bugs in the time series benchmark:

• ensure that item_id column is parsed as a string. If the column is parsed as an integer, it may lead to a different sorting order, which in turn introduces mistakes to seasonal error calculation.
• make sure that AutoGluon respects the random seed provided in the config
• add s3fs & fsspec dependencies that are necessary to enable pandas to load datasets from S3