This reorders the lfe_fir tables, and drops the mirrored half,
such that the loops in lfe_interpolation_fir() can be simplified.
The new loop structure should be easier to implement with SIMD.
Static data size is reduced by 2kB.
3% faster on Cortex-A8.

Originally committed as revision 22849 to svn://svn.ffmpeg.org/ffmpeg/trunk