hevcdsp: split the qpel functions by width instead of by the subpixel fraction

This should allow for more efficient SIMD.

Keep the C versions as they are now, to allow the compiler to inline the
interpolation coefficients.