Revert "Delete all alarms on abortAllDurableObjects()"

[kentonv]

Reverts #1918

Ironically, a week after this PR was merged, vitest-pool-workers actually removed its usage of this API. After two years of waiting...

And after some discussion, we feel that it doesn't make sense for abortAllDurableObjects() to cancel alarms. Cancelling alarms is essentially deleting data from the DOs, but since it doesn't delete all data, it leaves them in an inconsistent state.

It could make sense to add a new deleteAllDurableObjects() API which aborts DOs, cancels alarms, and deletes underlying data, but that should be a separate API.

[ask-bonk]

The revert is a clean, exact inverse of PR #1918. Every addition in #1918 is removed in #6355, and every removal in #1918 is restored. The changes are:

1. alarm-scheduler.h: Removes deleteAllAlarms() declaration and stmtDeleteAllAlarms prepared statement.
2. alarm-scheduler.c++: Removes deleteAllAlarms() implementation.
3. server.c++: Removes the alarmScheduler->deleteAllAlarms() call from abortAllActors().
4. Test files: Removes the AlarmObject class, test_abort_all_durable_objects_alarms test, related DO namespace/binding config, and reverts storage back to inMemory + drops rpc compat flag (both only needed for the alarm test).

No residual references to the removed functionality, no behavioral changes beyond the revert, no backward compatibility concerns (the reverted feature was internal/unsafe-module only). The PR description's rationale is sound — cancelling alarms without deleting all data leaves DOs in an inconsistent state.

LGTM

github run

[petebacondarwin]

I think there is still a need for deleting alarms in tests but something that can be worked on separately.

[kentonv]

Internal build flaked twice on a test that literally does not compile in any of the code modified in this PR, so I bypassed.