The range decoder used to initialize nbits_total after the
 renormalization loop, even though the renormalization loop
 modifies it.
This was presumably safe, because nothing actually used the value
 before it was initialized, but I'm tired of it triggering the
 integer overflow checking.