planner: prevent the optimizer from using different versions of stats of a table during optimization

## Enhancement
```
create table t (a int, b int, key(a), key(b));
insert into t values (1, 1);
analyze table t; -- stats-ver1
set global tidb_stats_load_sync_wait=1;

-- restart TiDB
explain select * from t where a=1; -- load stats-col-a

explain select * from t where a=1 and b=1; -- load stats-col-b timeout
+-------------------------+---------+-----------+---------------+-----------------------------------------------+
| id | estRows | task | access object | operator info |
+-------------------------+---------+-----------+---------------+-----------------------------------------------+
| TableReader_7 | 2.00 | root | | data:Selection_6 |
| └─Selection_6 | 2.00 | cop[tikv] | | eq(test.t.a, 1), eq(test.t.b, 1) |
| └─TableFullScan_5 | 2.00 | cop[tikv] | table:t | keep order:false, stats:partial[b:allEvicted] |
+-------------------------+---------+-----------+---------------+-----------------------------------------------+
```

In the second query, the optimizer uses 2 versions of statistics of table `t` where `stats-col-a` is loaded and then stats-ver1 is used while `stats-col-b` is not loaded and then pseudo is used, which seems risky.
Should we force the optimizer to always use one statistics version of a table during optimization? (all stats-ver1 or all stats-pseudo, no in-between state, not use them together).

Reproduce the SQL sequence in the issue, including the restart and tidb_stats_load_sync_wait setting, then trace the optimizer's table-statistics loading and fallback behavior. Determine how optimization should select one statistics version for all columns; done means the second EXPLAIN no longer combines loaded stats with pseudo statistics.

Written by the indexing model from the issue text.