Make 04811_prewhere_subcolumn_size_setting robust to randomized settings

The test compared the on-disk sizes of constant (highly compressible)
columns, so under the compression and granularity settings randomized by the
flaky check the compressed sizes were tiny and could invert, flipping the
PREWHERE condition ordering.

Use incompressible random strings with a large size separation
(exact(tup.small) < medium < size(tup)) so the ordering is stable regardless
of the randomized compression, granularity and serialization settings.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: Avogar

tests/queries/0_stateless/04811_prewhere_subcolumn_size_setting.sql

@@ -2,12 +2,16 @@
 -- `allow_calculating_subcolumns_sizes_for_merge_tree_reading` is enabled. When it is
 -- disabled, the subcolumn is costed by its whole top-level column size instead.
 --
--- The table has a Tuple where the `small` element is tiny and the `big` element is huge,
--- plus a `medium` column. With the setting on, `tup.small`'s exact size is smaller than
--- `medium`, so it goes first in PREWHERE. With the setting off, `tup.small` is costed by
--- the whole `tup` size (dominated by `big`), which is larger than `medium`, so `medium`
--- goes first. Both conditions use equality so both are "good"; statistics are disabled so
--- that only column sizes drive the ordering.
+-- The table has a Tuple where the `small` element is tiny and the `big` element is large,
+-- plus a `medium` column sized in between. With the setting on, `tup.small`'s exact size is
+-- smaller than `medium`, so it goes first in PREWHERE. With the setting off, `tup.small` is
+-- costed by the whole `tup` size (dominated by `big`), which is larger than `medium`, so
+-- `medium` goes first. Both conditions use equality so both are "good"; statistics are
+-- disabled so that only column sizes drive the ordering.
+--
+-- The data is incompressible random strings so that the size ordering
+-- exact(tup.small) < medium < size(tup) holds regardless of the compression and granularity
+-- settings randomized by the flaky check.
 --
 -- The analyzer path and the legacy InterpreterSelectQuery path are both covered; `tup.small`
 -- stays a subcolumn in both.
@@ -22,8 +26,8 @@ CREATE TABLE t_prewhere_subcolumn_size (id UInt64, medium String, tup Tuple(smal
 ENGINE = MergeTree ORDER BY id SETTINGS min_bytes_for_wide_part = 0;
 
 INSERT INTO t_prewhere_subcolumn_size
-SELECT number, repeat('m', 50), (repeat('s', 1), repeat('b', 500))
-FROM numbers(200000);
+SELECT number, randomString(30), (randomString(1), randomString(300))
+FROM numbers(100000);
 OPTIMIZE TABLE t_prewhere_subcolumn_size FINAL;
 
 SET enable_analyzer = 1;