We need this refactoring to move all node plan methodes to the new
Nodes structure. Otherwise we have to put a mutable lock for the
whole plan structure rather then on the nodes list. It makes
porblems with API refactoring otherwise.

BREAKING CHANGE: move from branch-based references in the tuple to
the array-based. Also add a logical ID to all relational nodes and
keep it in all its child tuples.

Now, our reference coordination system in a tuple looks like:
- parent relational node logical ID (we don't know actual position
  in the nodes list, so use "secondary index")
- target - i.e. a position of the reffered node in the relational
  node's children list
- posirtion in the reffered relational child node output tuple

Previously distribution was calculated in the node constructor. It
is a wrong design as the tree would be transformed multiple times
and there is no reason to recalculate distribution on each step.
We need distribution only on the last transformation when we insert
Motion nodes. So, this refactoring makes tuple distribution optional
and moves distribution logic to a separate function 'set_distribution()'.

We need it to serialize IR to SQL correctly - otherwise we'll lose
information about subquery name. Not good.

We expect that every relational node contains a reference to the
row expression node, that contains only aliases (expressions with
names). We need this convention to simplify SQL -> IR transformation.

Previously we had an alias only for reference, that was strange and
incorrect. Now we can add an alias name to any expression in the
tuple.

At the moment IR (logical plan) is far from final release. Current
commit fixes some point on a way to the final API and full test
coverage.

1. Implement columns (with types) and tables as a part of IR.
2. Use serde deserialization trait to build tables from YAML files.
   It would be also helpful when writing tests for arbitrary plan
   transformations.