Avoid redundant edge buffer copies in native IEF implementation

[rzumer]

Since IEF state is now propagated to the edge buffer initialization function, we can initialize larger edges for directional prediction with IEF and avoid cloning the edges in pred_directional() in case they need to be upsampled. In that case pred_directional() needs to process the edge in inverse order, since we use dav1d ordering for left edge samples, but it is a relatively simple optimization, might save a percent or so.