Split cwrsi() by pulses vs. dimensions.

This lets us cut out a bunch of work in the large _n, small _k case
 where most of the dimensions won't have any pulses.
It also gets rid of all remaining usage of CELT_PVQ_U() in cwrsi(),
 leaving just a single test instead of lots of mins and maxes, and
 makes a bunch of the jump threading more obvious.

This is a 1.6% decoder speedup on a 96 kbps comp48-stereo encode on
 a Cortex A8.