Keep the parent device (in this case the channel) around in ata_request,
so we dont panic device removal or failure.
Clean up ata_fail_requests to prevent the queue munging to fail.