Recently we've seen some oopses reported in code that cleaned
up HCs after they died ... like pci-pm not letting ohci-hcd
suspend until after the hardware was partly disabled, or users
removing a Cardbus adapter with ehci-hcd.  One root cause was
that the cleanup called hcd->stop() too many times.

This patch just does the cleanup that's reasonable to do
before the (dead) root hub is cleaned up:  it disconnects
the other devices.  And based on a suggestion from Pavel,
a diagnostic always appears -- avoiding mystery.