Currently the code initializes quantizer even
before reading it from bitstream and uses that
to dequantize samples.

The current range value causes an underflow
when negated and pushes anything less than zero
to the minimum.

For now force everything to int16_t, except
for variables used as index

If everything works fine for 16-bit streams,
the current code could be extended to do 24-bit.

Not sure if this is correct. But looks like
it was giving us off-by-one's when used to index
buffers.

(Resolves some conflicts)

Size should be determined from bits per sample.
The current code only accepts 16-bit streams.