Never retry tearDownAll (#2815)

A retried `tearDownAll` didn't run any of its callbacks again.
`Invoker.runTearDowns` removes each callback from the declarer's list
before running it, so by the time a retry started the list was empty. The
retry always passed, which hid the original failure.

Give the synthetic `(tearDownAll)` test `retry: 0` so a failure fails the
run, whatever `retry` the group, suite, or tags configure.
Platform-specific metadata, such as `on_platform` in `dart_test.yaml`, is
merged into the test's metadata later by `forPlatform`, so clear the retry
in those entries too. `setUpAll` still retries.

We chose not to make `tearDownAll` retries work. Re-running cleanup
correctly means tracking which callbacks failed, including ones that time
out or fail asynchronously, and the complexity cost is higher than the
utility. Teardown callbacks are often not safe to run twice anyway. A
`tearDownAll` callback that can fail on flaky cleanup should handle that
failure itself, for example by retrying within the callback.

This only surfaces failures that a retry was already hiding.

Fixes #2800