When forming a scaled prediction, we need to clamp against
the extended frame border which was set up when the relevant
reference frame was decoded. The width of this border actually
depends on the subsampling mode (for UV planes), but before this
patch we were always using the Y plane's border width.

This resulted in bad predictions when signalling a motion vector
which points far outside the reference frame. This patch fixes
the clamping, and restores the intended behaviour for out-of-frame
motion vectors.

Change-Id: I2cf575ce339a3e22a3c8444de0d0c3be031007c9