AAAARGH.  OK, I've got it to work with the current tree, but...  I don't
like the lifetime rules - both old and new ones.  Summary:

	* we create a gendisk upon attempt to open or upon successful
	  attempt to run (which _can_ happen without open).
	* we never remove them until the rmmod (or ever).
	* mddev is held by gendisk.

The source of trouble is our API - ioctls on possibly unconfigured
devices when no IO is possible, potentially setting up this (or another,
or many other) devices.  I _think_ we can get the old rules (mddev
exists when somebody holds it open or when it runs or when somebody is
looking at it funny), but the way I see will take more messing with
driverfs objects than I want right now. 

Notice that current API is, was and will be ripe with interesting races
- it really ought to be replaced with something saner, but that would
take serious modifications of userland tools.  I'll talk with Neil about
possible solution, this patch just gets it back to life.