There is a bug where we return from _dispatchOps() on shutdown-related failures, but _run() continues consuming batches. We should propagate the shutdown result back to _run() and stop the dispatch loop instead.

In addition to fixing the bug, this ticket should also add regression coverage that fails only the first inline op with a shutdown error and verifies that subsequent work is not attempted and lastApplied does not advance past the failed op. Also cover the failed-drain path.