This is an updated version of a patch submitted to me from
Michal Sojka <sojkam1@fel.cvut.cz>, basically providing a
much-needed rewrite of the highspeed ISO support.  I updated
the scheduling and made it a closer match to how OHCI works;
and also tested it a bunch.

So far it seems most of the requests for highspeed ISO support
have been for realtime data collection -- custom apps, nothing
a mainstream kernel would ship with.   The USB Video class is
now defined; highspeed video will also need these updates.

Key changes:

   - Define an "iso_stream" head for iso endpoints.  This acts
     enough like a QH that endpoint_disable() works.  It holds the
     queue of ITDs, and the endpoint's current schedule state.
     And it's easy to find (spinlocked array access), no search.

   - Uses a temporary "itd_sched" while submitting each URB, with
     not-yet-linked ITDs and request-specific metadata.  There's
     a per-stream cache of ITDs, so resubmitting ISO urbs (to
     achieve a "ring" of transfers) is typically cheap.

   - Scheduling for most URBs is almost a NOP:  just a sanity
     check to make sure there's no need to reschedule, and then
     just link into the schedule at the current schedule slot.
     (The previous code was a gross hack that didn't even work
     reasonably with more than two URBs queued.)

   - Is a reasonable model to use with full speed ISO transfers.
     (They need additional TT scheduling hooks, most of which
     are already written but not merged.)

   - Handles several cases the previous code didn't, including
     high bandwidth transfers (loads up to 24 MByte/sec)

   - Has had more testing than the old code, including 20+ hour
     successful IN+OUT runs, more varied transfer intervals and
     maxpacket sizes.  (Using net2280 and a gadgetfs driver.)

So it's worth replacing the existing code with this; there
aren't too many rough edges, and it's much more fixable than
the previous version.


p.s. Many thanks, Michal!