drm/xe/exec: A couple of fixes around the dma-fence signalling critical path

The first patch moves xe sync object dependency adding to before the dma-fence signalling critical path and also removes one point of failure from it.

The second patch deals with performing the userptr invalidation check before arming the sched job. That also moves away one point of failure from the dma-fence critical path and also a potential lockdep inversion since we want to be able to wait for fence inside the userptr notifier lock if needed (amdgpu does this, but IMO unnecessarily).

This is a rebase and fixup of !222 (closed) with the first patch skipped.