Add a call into the disk layer to synchronize the AIO queue. Wait for all
pending requests to complete. This will be necessary when resetting a
virtqueue.

The wait() operation isn't the same as flush(). A VIRTIO_BLK_T_FLUSH
request ensures that any write request *that completed before the FLUSH is
sent* is committed to permanent storage (e.g. written back from a write
cache). But it doesn't do anything for requests that are still pending
when the FLUSH is sent.

Avoid introducing a mutex on the io_submit() and io_getevents() paths,
because it can lead to 30% throughput drop on heavy FIO jobs. Instead
manage an inflight counter using compare-and-swap operations, which is
simple enough as the caller doesn't submit new requests while it waits for
the AIO queue to drain. The __sync_fetch_and_* operations are a bit rough
since they use full barriers, but that didn't seem to introduce a
performance regression.

Signed-off-by: Jean-Philippe Brucker <jean-philippe.brucker@arm....>