Further performance optimization of the atrac3 decoder. Rework the internal sample...

Further performance optimization of the atrac3 decoder. Rework the internal sample representation and usage of dsp routines. For now a quick and dirty solution is used to add a fract part of 2 bits. Through this several buffers and functions as well as copy loops could be removed. Furthermore add some ASM for coldfire and place some additional data in IRAM on PP5022/24 and X5/M5. Speedup on ARM: +3%, speedup on Coldfire: +639%. Both ARM and Coldfire can decode in realtime now.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@22561 a1c6a512-1295-4272-9138-f99709370657