The `online' modification of unsatisfied eqn counts caused increased rate of
decoding failures (verified experimentally). Use the variant that doesn't
modify the counts until next round.
This replaces the periodic recalculation of error correlations and the syndrome
by in-place modification. Bit flip is therefore a bit slower, but overall
decoding of the 256-bit secure variant fits in 200ms, and 128-bit variant
decodes under 20ms.
There still could be some (blatantly nondeterministic) method to do this using
FFT, research underway.