[Intel-gfx] Updated fence patches

[Jesse Barnes]

This set includes a sketch of how we might allow fences to be emitted
directly within a batch buffer.  This gets rid of the need for flushing
around fence operations, which can be a win, and lets userspace more
finely control things.

If it looks reasonable, we could drop the separate ioctl and just use
this instead...  Alternately, I could still add the "return a fence from
execbuf" behavior to the main execbuf ioctl, to represent completion for
the whole thing.  And per the last comments, we still want a way to pass
an array of fences in with a given operation, so it can be properly
ordered by the scheduler.

Jesse