Fix LDPC decoding accuracy

A poster on our forum observed that the accuracy of our LDPC decoder
was very much worse than that of other implementations.

The cause is that we do all our calculations in int8_t fixed-point,
which quickly saturates. The fix is to do all intermediate
calculations in int16_t fixed-point.

On average this causes a performance regression of between 1.2x and
1.3x across all ArmRAL benchmarks.