Write basic feature stats to tables after feature generation

[thcrock]

@saleiro @ecsalomon and I went on a bit of a dive this morning following my exploration of the https://github.com/tdda/tdda repository . The team has talked about "premodeling" before, that could expand in scope to be as big as postmodeling. Right now, however, I think we should figure out whats the simplest possible thing that could be useful as a beachhead for this premodeling. And in our discussion I think we have something.

After feature generation (both pre-imputation and post-imputation), create feature stats in table/s that encompass basic min/max/nulls stats for the created features. These would always reflect the current state of the entire feature table (e.g. we recompute each time), provided that this doesn't take too long.

[saleiro]

data_dict_df = pd.DataFrame()
for index,row in cleaned_cols_df.iterrows():
    #this might take a while, let's print some stuff while it runs
    if row['data_type'] != 'USER-DEFINED':                                          #FIXME - error with user-defined
        print(row['table_name'],'--', row['column_name'])
        data_dict_df = data_dict_df.append(
        sql("""
            select '{table}' as table_name, 
                    '{col}' as column_name, 
                    '{dtype}' as data_type,
                    count(*) as count_rows,
                    count(case when {col} is null then 1 end) as count_nulls,
                    count(distinct {col}) as count_distinct, 
                    min(case when pg_typeof({col}) in ({num_datetime}) then {col} else null end),
                    max(case when pg_typeof({col}) in ({num_datetime}) then {col} else null end),
                    avg(case when pg_typeof({col}) in ({num_datetime}) then {col} else null end),
                    sttdev(case when pg_typeof({col}) in ({num_datetime}) then {col} else null end),
                    mode() within group (order by {col})
            from cleaned.{table}
        """.format(table=row['table_name'], col=row['column_name'], dtype=row['data_type'], num_datetime=','.join(num_datetime_types)))
        ,ignore_index=True)

For feature tables, since all columns are non-categorical there is no need to verify the pg_typeof.