* don't synchronize_irq() in cp_stop_hw().  when necessary, do so
  in the callers.
* call synchronize_irq() in cp_close(), after releasing the spinlock.
* move netif_stop_queue() and netif_carrier_off() calls inside lock
* flush last interrupt status write, in cp_stop_hw()
* add unlikely() check for dev==NULL, first thing in the irq handler