[YCQL] Index scan and parallel scan need to use a consistent read time when doing a multi page scan

[lnguyen-yugabyte]

Jira Link: DB-3098

Description

Currently for index scan and parallel scan, a consistent read point is not enforced. That results in an index scan having different read time on each page, which might cause one to read a partial update of a transaction (similar to the issue in #10328).

Example

CREATE TABLE t (k int primary key, v1 int, v2 int) WITH transactions = "
      "{ 'enabled' : true };

INSERT INTO t (k, v1, v2) VALUES (0, 0, 20000);
INSERT INTO t (k, v1, v2) VALUES (1, 0, 0);

CREATE INDEX idx ON t(v1);

Now on two ycqlsh shells, execute the following concurrently:

ycqlsh1>   BEGIN TRANSACTION
                   ... update t set v2 = v2 - 10 where k = 0;
                   ... update t set v2 = v2 - 10 where k = 1;
                   ... insert into t (k, v1, v2) values (2, 0, 10) if not exists else error;
                   ... insert into t (k, v1, v2) values (3, 0, 10) if not exists else error;
                   ... END TRANSACTION;

ycqlsh2> paging 1;

ycqlsh2>
explain select (k, v1, v2) from t where v1 = 0;

 QUERY PLAN
-----------------------------------------------------------
 Index Scan using ybdemo_keyspace.idx on ybdemo_keyspace.t
   Key Conditions: (v1 = 0)

ycqlsh2> select k, v1, v2 from t where v1 = 0;

The result would be

 k | v1 | v2
---+----+-------
 0 |  0 | 20000

---MORE---
 k | v1 | v2
---+----+-----
 1 |  0 | -10

---MORE---
 k | v1 | v2
---+----+----
 2 |  0 | 10

---MORE---
 k | v1 | v2
---+----+----
 3 |  0 | 10

which is wrong.