Also minor C89 fix for the previous commit

We only keep when concealing less than 10ms with SILK.

Since we're doing two rotations, we can invert the sign on both.
Also adding a few comments for optimizing the FFT.

Also simplifying the code

Undoes commits f7547a4e and 72513f3c

This also happens to increase the accuracy since it appears that the new
ordering is optimal (at least for 20 ms frames), whereas the previous ordering
was pessimal.

IMDCT now uses the forward FFT.

This makes us lose a bit of precision in the first steps, but our gain is more
precise because it's only rounded once. Overall, SNR is slightly improved.

Also moves the VSHR32() condition outside the loop just in case some compilers
don't optimize that properly.

This reduces the peak decoding error by removing small (inaudible) spikes in
the error at the frame boundaries. These were due to the frequency-domain bias
ending up as a small pulse in the middle of the IMDCT overlap. None of this
was ever audible, but fixing it is still cleaner.