Removing redundant calls to memcpy from
build_wedge_inter_predictor_from_buf yields a net 4% encoder speedup
with ext-inter only. The output is identical.

Change-Id: If97d4e323a5c8aca90c84a25a72085e006b05446