When a gadget is disabled, kill_all_requests() can be called
simultaneously from both a user process via dwc2_hsotg_pullup() and from
the interrupt handler if the hardware detects disconnection.

Since we drop the lock in dwc2_hsotg_complete_request() in order to call
the completion handler, this means that the list is modified
concurrently and leads to an infinite loop in kill_all_requests().

Replace the foreach loop with a while-not-empty loop in order to remove
the danger of this concurrent modification.

Note: I observed this with threadirqs, I'm not sure if it can be
triggered without threaded interrupts.
Signed-off-by: John Keeping <john@metanate.com>
Signed-off-by: Felipe Balbi <felipe.balbi@linux.intel.com>