Fix "DB::Exception: Block structure mismatch in UNION stream: different number of columns"

[chenxing-xc]

I hereby agree to the terms of the CLA available at: https://yandex.ru/legal/cla/?lang=en

I encounter the following failure, and think it's a bug. The version I am using is stable 1.1.54380, and could also reproduce this problem in main branch.

== How to reproduce:

• Setup a cluster with more than one shards
• create mergetree table and distributed table
  CREATE TABLE default.t ( id Int64, p_date Date) ENGINE = MergeTree PARTITION BY p_date ORDER BY id SETTINGS index_granularity = 8192
  CREATE TABLE default.t_all ( id Int64, p_date Date) ENGINE = Distributed(xxx, default, t, intHash64(id))
• client connect to one of shard in the cluster, and issue the following statement, and could see the problem:
  ------- :) select id from t_all where p_date='2018-01-01' limit 1

SELECT id
FROM t_all
WHERE p_date = '2018-01-01'
LIMIT 1

Received exception from server (version 1.1.54382):
Code: 171. DB::Exception: Received from localhost:9000, ::1. DB::Exception: Block structure mismatch in UNION stream: different number of columns:
id Int64 Int64(size = 0)
id Int64 Int64(size = 0), p_date Date UInt16(size = 0).

== Problem analysis:
After predicate is moved to prewhere expression, the rule to deduce columns in the result stream should be adjusted, otherwise, we should keep the PROJECT action generated by the original prewhere expression which help deduce correct stream header(block's columns).

[zhang2014]

#2226 seems like a more complete solution.