Here's a streamlined version:  doesn't require the unlink
cleanup, and expects the *hci-hcd updates later.  Smaller,
and sanity checked against all three major HCDs.

  - reverts that unlink() patch: disconnect() callbacks
    can continue to act like they always have.

  - adds new "disable that endpoint" support, necessary
    for safely changing configurations or altsettings
    as well as physical disconnect (which is almost the
    same as setting config to zero, except for ep0).

  - NEW BEHAVIOUR:  usbcore cleans up after drivers that
    return from disconnect() with urbs still linked, by
    using the new "disable that endpoint" support.

Because it doesn't have any *hci-hcd updates, the hardware
synch needed by EHCI and OHCI (not UHCI) still gets done
through the bus->deallocate(dev) call ... not where we've
ever needed it, but 2.3-compatible (and finally fixable).

That gets rid of some problematic disconnect scenarios, and
other fixes can be phased in over time:  the *hci-hcd updates
to relocate the hardware synch point so it always happens in
a task context, and disabling endpoints before their configs
change (invalidating all HC state).