join filter with separate input rows instead of merge them first

[zhouqingqing]

This is what we have now for join filter. We first merge (l, r) then do the filtering as the filter.Exec() API only takes one row as input.

                            Row n = new Row(l, r);
                            if (filter is null || filter.Exec(context, n) is true)

we may want it becomes

                            if (filter is null || filter.Exec(context, l, r) is true){
                                 Row n = new Row(l, r);

which is more efficient. To do so:

1. extend Expr.Exec(context, Row l[2]) with only BinaryExpr supporting it.
2. extend ColExpr to have a "int joinordinal_" which takes values 0 or 1 indicating it is "l[0]" or "l[1]".

The problem of above method is that every expr get affected a bit performance wise. Meanwhile, if we want to support nary join directly, we can extend Row l[2] to list?