test(zebrad): the serialized live-mainnet tests set a 10-minute floor on the unit-test job

[upbqdn]

Measured on a 32-core host, the ci nextest profile runs 1320 tests in 623s wall while the tests only sum to 34 minutes of work, because the serial-live-mainnet group has max-threads = 1 and its members sum to ~10.3 minutes — the whole wall time.

Those members spend the time sleeping, not working. Durations land on the constants in zebrad/tests/common/launch.rs: EXTENDED_LAUNCH_DELAY (45s) for the four unit::config::*ephemeral* tests, LAUNCH_DELAY/BETWEEN_NODES_DELAY (20s each) for unit::cli::start_*, the integration::network::zebra_*_conflict pair-launch tests, integration::database::zebra_state_conflict, and the integration::rpc::*_endpoint* tests.

They are in the serial group because they start a node with the default DNS seeders, so concurrent runs from one IP reject each other handshakes.

Two directions, either of which cuts the job to about a minute: wait on the log line that proves the node reached the state under test instead of a blanket delay, or run these tests without live seeders (they do not test peer connectivity) so they leave the serial group and run concurrently.

[conradoplg]

Agreed, we should do one of these depending on what makes sense for each test:

• Do not wait an arbitrary amount of seconds, but rather loop waiting for whatever condition we need with a max timeout
• Replace the live mainnet test with a regtest node
• Move it to a daily/weekly integration test run if is is truly a useful network test that needs mainnet