nvme_driver: restored driver hangs on shutdown due to stale admin queue

After NvmeDriver::restore() in a unit test, calling shutdown() on the restored driver hangs forever.

Root cause

The admin queue always has a pending AER (Asynchronous Event Request) command. During restore, DMA blocks are freshly allocated at different physical addresses than the originals. The NVMe controller's internal SQ/CQ BAR addresses still point to the old blocks, so the AER completion never arrives. shutdown() waits for all pending commands to drain, which never happens.

What was tried

• std::mem::forget(driver) — leaks but avoids the hang; not a real solution
• update_servicing_flags(false) + shutdown() — triggers a controller reset, but the reset path itself waits for pending admin commands
• Clearing nvme_keepalive and calling shutdown() — same hang

Possible fixes

• Make the AER task cancelable so shutdown can kill it without waiting for a completion
• Add a test-only shutdown_force() that drops queues without draining
• Fix the DMA identity problem so the controller sees completions (depends on the PagePool blocker in #489)

Start with the unit test that calls NvmeDriver::restore() followed by shutdown(), then trace the restored driver's admin queue and pending AER handling. Read the restore and shutdown entry points and consider the PagePool blocker in #489; done means shutdown completes without waiting forever for a stale AER completion.