[Intel-gfx] Demidlayer i915 loading

[Chris Wilson]

Now that the changes to allow reordering of connector registration have
landed in drm-misc, we can look at proceeding with removing the
dev->driver->load() callback and take control of our initialisation
sequence, with the goal of not exposing driver internals to userspace
before we are ready.

(As with previous series, my ulterior motive is to close a hole in GEM
context vs RPS initialisation which depends upon closing the userspace
hole, and so fix a regression whereby we try to initialise a context
during suspend.)
-Chris