[matteobovetti]

What

The test "with the database down, the first read waits the acquire timeout" asserted that the
message contained the driver's word refused. That holds only if the driver's connect has already
failed when the 500 ms acquire timeout fires. Under load it had not, and the pool correctly said
"none recorded yet".

How

• The first borrow now asserts only what does not depend on that ordering: the wait, the pool's
  own words, and that neither the password nor detached leaks into the message.
• A new test asserts refused. It borrows until the pool has the driver's failure on record, with
  a 10 s deadline, so a real regression still fails with the last message.
• bin/gate no longer lists PoolSuite as timing sensitive.

Testcontainers 1.21.4

Docker Engine 29 answers testcontainers 1.21.3 with an empty 400, so no suite that starts Postgres
could run. Bumped to 1.21.4 in project/Dependencies.scala. The examples build kept its own
literal 1.21.3 and stayed red, so blog and shop now take Dependencies.testcontainersPg through
the examples meta build, and the version has one home.

Not in this PR

ConnectionUnavailableSuite has the same race at a 250 ms timeout. Left for a follow up.

Closes #320