tests: fix test flakiness by avoiding localhost and adding port retry

[tbutter]

The flakiness came up during nix package upgrade NixOS/nixpkgs#513622 (comment)

AI disclosure: Used antigravity cli to find the flaky test and create a patch.

1. Host/DNS Resolution Flakiness ( localhost vs 127.0.0.1 ): The tests previously attempted to verify port reachability and perform test requests using localhost . In sandboxed build environments (such as the Nix sandbox), hostname resolution for localhost can
   be unreliable or resolve to the IPv6 loopback address ( ::1 ) first. Because dummyhttp binds to 0.0.0.0 (IPv4-only by default), connecting via IPv6 fails or times out.
2. Parallel Port Allocation Race Conditions: The test suite uses the port_check crate to query for a free port, which it then closes before starting the dummyhttp process. Since tests run in parallel, multiple tests could check for a free port, get the same port
   number, and cause dummyhttp to fail to bind on startup. Additionally, the startup loop would wait for the full timeout (1 second) instead of immediately failing-fast if the child process exited early.

[svenstaro]

Hey, thanks. I think a better way would be to adapt what we do over in buildbtw. That way, we can ensure a process will definitely get the port that it's locking because we can pass the lock around.

[svenstaro]

Also I'd appreciate it if you could separate out the DNS stuff to a second commit.