test(migrations): Deflake test_reverse_idempotency_all

[phacops]

test_reverse_idempotency_all snapshots system.tables before and after a reverse/fake-forward cycle and compares the two results for equality. The snapshot query had no ORDER BY, and system.tables has no guaranteed row order, so the comparison was order-sensitive.

In the distributed setup the reverse_twice() cycle between the two snapshots issues ON CLUSTER DDL (CREATE … IF NOT EXISTS / DROP … IF EXISTS) which can re-register tables and change the order rows come back in without changing the set of tables. That produced intermittent failures in the test_migrations_distributed CI job — e.g. on #7980, whose latest EAP migration drops and recreates three downsample materialized views on reverse, which is exactly the kind of churn that perturbs the row order. The reverse cycle itself is a no-op on the table set (fake-forward only updates status; the create/drop are guarded by IF [NOT] EXISTS), so the failures were spurious, not real state differences.

This orders both snapshots by create_table_query so the assertion reflects the set of tables and their definitions rather than the iteration order. A genuine difference in tables or definitions is still caught.

Also replaces exc.value.args[0] with str(exc.value) in test_check_inactive_replica — a pre-existing mypy error (BaseExcT_co_default has no attribute "args") in this file that pre-commit only surfaces once the file is otherwise modified. str(exc.value) is the equivalent message check and types cleanly.

Agent transcript: https://claudescope.sentry.dev/share/5BgPWGjZzfAo_gfbl1kqFzjsOJKf5sz8T4u3s0en9OY