Foreign keys should be mapped using columns<->ref_columns instead of columns<->primary_key

[neumannjan]

deep-db-learning/db_transformer/data/dataset.py

Lines 270 to 279 in 393c6e2

	table_primary_keys = primary_keys[col.ref_table]
	for i_index, row in enumerate(table_data):
	if row[index] is None or row[index] not in table_primary_keys:
	continue
	edge_index.append([i_index, table_primary_keys[row[index]]])
	hetero_data[table_name, col.columns[0], col.ref_table].edge_index = (
	torch.tensor(edge_index, dtype=torch.long).t().contiguous()
	)

As the title of this issue suggests, the proper way to map foreign keys should not be based on the primary key, since most SQL databases allow more general foreign key constraints that do not require the referenced set of columns to be the primary key.

Foreign key constraints in the DB (and consequently also those in our schema) ensure that len(columns) == len(ref_columns).

This should also eliminate the limitation of our code that requires a foreign key constraint to consist only of a single column.