ALTER TABLE will refresh previous snapshot values for absent cells

[BugenZhao]

There's a field named snapshot_value in DefaultColumnDesc, which is the evaluation result of the default expression at the time when it's created or added with ALTER TABLE. This is introduced in #12678 to support impure default expressions like DEFAULT NOW().

If we then call ALTER TABLE again, the table will be replanned and the snapshot values will be evaluated once more. The new values could be different from the previous one if the expression is impure, leading to inconsistent select results for pre-existing rows prior to the very first ADD COLUMN .. DEFAULT NOW().

To reproduce:

dev=> create table t(v int);
CREATE_TABLE
dev=> insert into t values (1);
INSERT 0 1
dev=> alter table t add column now1 timestamptz default now();
ALTER_TABLE
dev=> select * from t;
 v |             now1              
---+-------------------------------
 1 | 2024-06-05 05:50:41.723+00:00  <- snapshot value
(1 row)

dev=> insert into t values (2);
INSERT 0 1
dev=> select * from t;
 v |             now1              
---+-------------------------------
 1 | 2024-06-05 05:50:41.723+00:00. <- snapshot value
 2 | 2024-06-05 05:50:56.326+00:00. <- real-time value
(2 rows)

dev=> alter table t add column v2 varchar;
ALTER_TABLE
dev=> select * from t order by v;
 v |             now1              | v2 
---+-------------------------------+----
 1 | 2024-06-05 05:51:43.370+00:00 |     <- snapshot value got updated unexpectedly!
 2 | 2024-06-05 05:50:56.326+00:00 | 
(2 rows)

In Postgres, since the table will be rewritten to fill the default values for pre-existing rows, there's no such issue. We should be consistent with that behavior.