Some small updates:

  - Sometimes read requests can be satisfied directly from
    the OUT fifo.  This fixes a bug where the return code
    from usb_ep_queue() overwrite the transfer status, which
    in that case was set _before_ that call returned.
    (Synchronous behavior; not the usual async completion.)

  - In the same vein,  usb_ep_dequeue() doesn't need to be
    synchronous -- though so far most controller drivers
    have implemented it that way.  So drop the spinlock
    before the wait_event() sleep.

  - Some debug messages are more useful AFTER the event
    than before.

  - The only descriptor fetches user mode drivers will
    need to handle are for string descriptors.  Stall all
    other requests, like ones for other-speed configs
    on single-speed devices.