Rockbox

This patch speeds up the rolo_load function, by
implementing the descrambling part in assembler (and in
a more efficient form). Also, the descrambling
calculates the checksum as it goes, saving a little
more time. Descrambling time is reduced from about 1.6
to 0.3 seconds (on a player, with a .mod file on
122,534 bytes).

As the patch also moves the "critical" part of
rolo_load to the internal RAM (this, including
descrambling and checksumming, needs 88 bytes), the
.topcode section is removed.

descramble.S is attached separately, because I don't
know how to make cvs diff include it in the patch… :/