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.
