platform-checks: fix silently ignored flags, replica masking, state leaks

[def-]

• Plumb --system-param and --default-replication-factor through to the Mz instances started by StartMz. Both flags were silently ignored because StartMz overrides the service definition wholesale. Also parse all values of a multi-value --system-param instead of only the first.
• UseClusterdCompute: drop all quickstart replicas, not just r1. Since the default replication factor is 2, the orchestrated r2 kept serving queries and masked compute restarts in RestartClusterdCompute and the UpgradeClusterdCompute* scenarios.
• cloudtest: sort checks before sharding so that parallel jobs agree on the shard split. Set iteration order differs between processes, so checks could run twice or not at all.
• Wipe the environment between scenario runs (and between checks in oneatatime mode), iterate scenarios in deterministic order, use a fresh executor per run, and copy the check list before shuffling. Previously state leaked from one run into the next.
• Make Executor.system_settings per instance and record settings as applied only after the testdrive applying them has succeeded.
• Backup scenarios: restart Mz through StartMz instead of Restore(restart_mz=True) so the restored instance keeps the service definition it was started with (e.g. the SASL listener).

Test runs: https://buildkite.com/materialize/nightly/builds/17084 & https://buildkite.com/materialize/release-qualification/builds/1303