Fixup and improve per-thread data

Fixup and improve per-thread data

- Get rid of backend_data, just use sk_out
- Add reference counting to sk_out and proper assign/drop helpers
- Ensure we pass it to other threads created
- Fix leaks of sk_out

This should get us closer to the new model actually working as
intended.

Signed-off-by: Jens Axboe <axboe@fb.com>