Isolate dataset for qualx plugin invocations

[leewyang]

This PR isolates the rows for the targeted dataset when invoking a qualx dataset-specific plugin.

Previously, all plugins were invoked with the entire input dataframe, which relied on the plugin implementation to correctly target only the rows of its intended dataset. While this design was simple and allowed for the full flexibility to modify the data in any way, it also can lead to unexpected behavior if a plugin accidentally targets the rows of other datasets.

I have confirmed that the full workflow of preprocessing, training, and evaluate produces the same models and evaluation results.

Changes

1. send only the slice of the input dataframe for the specific dataset targeted by the plugin.
2. raise an error if the plugin code modifies the dataframe indices (implying invalid/unsupported modification of the rows).
3. add support for a 'default' split_function (by name vs. by presence in a list).
4. updated qualx.md doc with more details for plugin implementations.

Test

Following CMDs have been tested:

Internal Usage:

python qualx_main.py preprocess
python qualx_main.py train
python qualx_main.py evaluate

[parthosa]

LGTME. @eordentlich Could you please review it as well?

[eordentlich]

LGTM 👍