Use fdct8x4 instead of fdct4x4 where the block size allows it.

This allows for faster SIMD implementations in the future (currently
there is no speed impact).

Change-Id: I732647e9148b5dcb44e6bc8728138f0141218329