Resolve "Missing energy on stereo CNG frames at 24k4 kbps" [allow regression]

Closes #732 (closed)

There is a mechanism to prevent values below 1e-5 in the msPeriodog calculation in fd_cng_dec_fx.c. However, the adaptive Q-value may lead to this value being zero. This patch uses the smallest representable non-zero value instead, meaning 1 in the used Q-value.

An alternative fix, that also solves the issue with the spectral hole, is to use the EVS version of the function perform_noise_estimation_dec_fx instead of perform_noise_estimation_dec_ivas_fx. It seems to handle the signal dynamics in a better way. However, there is a side effect that a HB noise burst appears later in the signal. I think the preferred solution would be to use the EVS variant of the function, and to find out the reason of the new noise burst.