Add integration tests for PgmqTransport against a real pgmq

[iGrog]

Refs #2, #3.

The PGMQ transport had no tests. This PR adds:

• a pgmq service to compose.yaml (ghcr.io/pgmq/pg18-pgmq:v1.11.1, the same image thesis/pgmq uses) and THESIS_PGMQ_DSN for the php service. make test already runs up, so the CI needs no changes;
• PgmqTransportTest::consumerDrainsBacklogWithoutWaitingForPollInterval;
• PgmqNotifyTriggerTest:
  • consumerStartDoesNotWaitForProducerTransactions: notifications set up, a producer keeps an insert open, the start must not wait;
  • concurrentConsumerStartsCreateTheTriggerOnce: 4 workers (separate pools, like worker processes) start at once on a queue without notifications, while an open producer transaction forces the starts to overlap; CREATE TRIGGER is counted with an event trigger and must be exactly 1, with no failed starts;
  • consumerStartRestoresNotificationsLostInCrashRecovery: the queue's row in the UNLOGGED pgmq.notify_insert_throttle is removed (what crash recovery does), and a start must bring notifications back;
  • consumerStartRestoresDisabledTrigger;
  • createdQueueWakesConsumerOnInsert: createQueue() sets notifications up.
• the one-line phpstan fix from Remove the @phpstan-ignore that no longer matches #5, so the run shows only the tests' own result.

Whether the package's dev environment should include Postgres is your call. That is why the fixes (#6, #7) are separate PRs that do not depend on this one.

Expected CI result

• On its own (on 0.5.x): exactly 3 tests fail, on purpose, one per defect: consumerDrainsBacklogWithoutWaitingForPollInterval (PgmqTransport: consumer reads one batch per poll signal, a backlog waits pollInterval between batches #2), and consumerStartDoesNotWaitForProducerTransactions and concurrentConsumerStartsCreateTheTriggerOnce (PgmqTransport: every consumer start recreates the notify trigger (table lock, blocked producers, deadlocks) #3). The concurrent test collects start failures instead of leaking them into other tests.
• With Pgmq: poll again right after a full batch instead of waiting for the next signal #6, Pgmq: ensure insert notifications in createQueue() and on consumer start (via thesis/pgmq) #7 and a thesis/pgmq containing Do not recreate the notify trigger on consumer start; ensure notifications under the queue lock pgmq#22: 14/14 green (verified locally against that pgmq branch).

The event-trigger test needs a superuser connection, which the compose service provides.

[iGrog]

Updated the tests:

• new concurrentConsumerStartsCreateTheTriggerOnce (event trigger counts CREATE TRIGGER; 4 workers with separate pools; collects start failures so they don't leak into other tests);
• new consumerStartRestoresNotificationsLostInCrashRecovery and consumerStartRestoresDisabledTrigger;
• the phpstan fix from Remove the @phpstan-ignore that no longer matches #5.

On 0.5.x exactly 3 tests fail, one per defect. With #6, #7 and thesis-php/pgmq#22 all 14 pass (verified locally).