Replace split_nds with split_train_val

[leewyang]

This PR replaces the split_nds function with a more generic split_train_val function. This avoids having any dataset-specific split functions in the core qualx code, and pushes them to the dataset-specific plugins.

I have confirmed that this produces identical models when paired w/ a new nds.py plugin in the internal repo.

Test

Following CMDs have been tested.

Internal Usage:

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

[parthosa]

Thanks @leewyang. LGTM. Small question:

What does this refer to?

pushes them to the dataset-specific plugins

[leewyang]

@parthosa basically, some datasets require custom logic for determining train/test splits, e.g. we use a fixed/curated list of queries for NDS. This logic is now moved to a plugin that gets loaded per this code.