drm/sysfs lifetime interaction fixes

[Dave Airlie]

As reported in the thread
[xen] double fault: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC
on Linux kernel the drm has a crappy interaction model with the sysfs objects.

This is my preferred method of fixing it as I don't think the lifetimes need
to be tied so closely, though this requires review my someone to make sure
my unregistering etc is correct and in the right place.

Russell helped my understanding of the device model and supplied a patch
doing this another way, but I think I like splitting the lifetimes more.

Patch 1 is just prepartory for i915, then patch 2 is the guts.

Dave.