There are several weirdness at the unregister logic.

First of all, IR has a poll thread. This thread needs to be
removed, as it uses some resources associated to the main driver.
So, the driver needs to explicitly unregister the I2C client for
ir-kbd-i2c.

If, for some reason, the driver needs to wait for a close()
to happen, not all memories will be freed, because the free
logic were in the wrong place.

Also, v4l2_device_unregister() seems to be called too early,
as devices are still using it.

Finally, even with the device disconnected, there is one
USB function call that will still try to talk with it.
Signed-off-by: Mauro Carvalho Chehab <mchehab@redhat.com>