After a discussion here, we can rely on parquet to load data from OpenML.

Since we introduce the parser keyword, we can provide parser="parquet" that can rely on some optional dependencies to load the dataset.

In addition, we should check the consistency between parser="pandas" and parser="parquet".