make `COLUMNS OF <generator>' a first-class table in BQL

[riastradh-probcomp]

No description provided.

[raxraxraxraxrax]

Can you provide a possible use case for this, please?

[riastradh-probcomp]

Currently we have the two ad-hoc constructions

ESTIMATE ... FROM COLUMNS OF <generator>

and

ESTIMATE ... FROM PAIRWISE COLUMNS OF <generator>

Suppose we had three-way or n-way column functions, e.g. conditional mutual information. Then we'd have to add

ESTIMATE ... FROM TREBLEWISE COLUMNS OF <generator>

and so on -- but that's silly. All of these are naturally just selects on joins of the table of variables in the generator in question.

Not everything involves BQL functions. E.g., the function describe_generator_columns in bdbcontrib falls back to SQL and does a complex join on bayesdb_generator, bayesdb_column, and bayesdb_generator_column, which would be much better expressed via BQL using a select on COLUMNS OF <generator>.

You don't always want to join the columns of a generator with itself -- you might want to look up metadata about a column in some model by doing a join of COLUMNS OF <generator> and, say, bayesdb_crosscat_column or something. Or you might want to do it with multiple different tables, using qualified names:

ESTIMATE ... FROM COLUMNS OF <generator0> AS c0, COLUMNS OF <generator1> AS c1 ...