I.e. we can build queries like

select * from t1 join (select "COLUMN_3" as a, "COLUMN_4" as b from
(values (1, 2), (1, 3))) as t2 on t1.a = t2.a

At the moment it seems that we are enough with two distribution
types: by key (segment) and not by key (any). So, remove all other
types as redundant.

Added a test case with subquery as well.

Now we can traverse only relational nodes in the tree in arbitrary
order.

It is required to traverse the chain of "and" boolean expressions.

At the moment we don't add tests as upcoming equality propagation
would implement enough tests for this scenario.

We don't want to mess with subqueries in projection as our query
set doesn't contain such queries. It also simplifies some
transformations (equality propagation for example).