Make InsertData independent of block boundaries and order

[Hakaou]

Noticed while running the client suite under UBSan for #575.

InsertData inserts two blocks of four rows, then checks the selected rows with a branch on which block is being read, indexing TEST_DATA with the running row number. When the server returns the eight rows in a single block, the first branch runs for all eight rows and reads TEST_DATA[4..7], past the array:

client_ut.cpp:625:21: runtime error: index 4 out of bounds for type 'const struct (unnamed struct)[4]'
client_ut.cpp:626:21: runtime error: index 4 out of bounds ...
client_ut.cpp:627:21: runtime error: index 4 out of bounds ...

The reads happened to return the expected values, which nothing guarantees.

The select also has no ORDER BY. With insert squashing disabled (min_insert_block_size_rows = 0, min_insert_block_size_bytes = 0) and max_block_size = 1, a 25.12.13.2 server keeps the two inserted blocks separate and sometimes returns 2, 4, 6, 8 first: 3 of 600 runs of the current test failed that way locally. 26.11 stores them as one block, so it did not reproduce there.

To reproduce on 25.12: add the three settings to the default profile of ci/docker-compose/users.xml, start the server with it, then run ./clickhouse-cpp-ut --gtest_filter='*InsertData*' --gtest_repeat=300 on master.

Change: one Row struct, both datasets as vectors, the select orders by id, the expectation is both datasets sorted by id, and every received row is compared against it with ASSERT_LT(row, expected_rows.size()) instead of the branch. Block boundaries and block order no longer matter.

Checked locally (macOS arm64) against 25.12, 26.3 and 26.11: the test passes under AddressSanitizer and UBSan with halt_on_error=1, 600 runs in a row with the settings above on 25.12, and the whole ClientCase suite stays green.

[CLAassistant]

All committers have signed the CLA.