This should reduce computation, reduce stack usage, and be
substantially easier to read, but should not change behavior.
It's not quite bit-exact because I collapsed the application of the
decay and the fading into a single pass to save one multiply per
sample, but the difference signal is silent.
It also changes the comments into complete sentences.
We always define OPUS_BUILD, so this code has no effect.
Likewise there is no reason for it to be conditional.
According to Jean-Marc it's left over from when we had
an alternate CELT_BUILD define.
