refactoring: insert redistribution

Previosly we had a very complicated logic for INSERT operator:
we always generated a motion node under INSERT and generated
bucket_id with a special motion command in this virtual table.
As a side effect we made motions even for those INSERTs that was
possible to execute locally.

Another problem was caused by the column type derivation in
INSERT .. VALUES .. expressions. If any column in the first row
contained NULL it confuses Tarantool and it returns scalar (or
even boolean!) type instead. And there is no projection type cast
in Tarantool's insertion. As a result we often had type mismatch
errors in INSERT .. VALUES .. queries.

The solution is to rewrite this weird logic from the scratch.
Now we rewrite the tree somehow like this:

insert into t1 (a, b) values (1, 2)

=>

insert into t1 (a, b, bucket_id)
select
  COL_1,
  COL_2,
  bucket_id(
    coalesce(cast(COL_2 as string), '')
    || coalesce(cast(COL_1 as string), '')
  )
from (
  select
    cast(COLUMN_1 as integer) as COL_1,
    cast(COLUMN_2 as unsigned) as COL_2
  from (values (1, 2))
)