unrolls some loops, utilizing all 8 xmm registers. fft-test
shows ~10% speed up in (I)FFT and ~8% speed up in (I)MDCT on Dothan


git-svn-id: file:///var/local/repositories/ffmpeg/trunk@9017 9553f0bf-9b14-0410-a0b8-cfaf0461ba5b