Changed methode of mean power calculation.

[Mictronics]

I changed the method to calculate mean power in the converter. We already have the average mean level so we can calculate average power from it. This corrects also the mean power value because what is calculated by the old method is slightly off, probably due to loose in precision when casting.

Example from live IQ data:
sum_level = 91873866
out_mean_level = 0.010695526
Level to dB is: 20log(out_mean_level) = 20log(0.010695526) = -39.416 dB

Old method (summing up mag^2):
sum_power = 85686914156
out_mean_power = 0.00015221812
Power to dB is: 10log(out_mean_power) = 10log(0.00015221812) = -38.175 dB
Converting this power back to level results in 0.0123377 so it's not equal to square level (or magnitude).

New method:
output_mean_power = out_mean_level^2 = 0.000114394
Power to dB is: 10log(out_mean_power) = 10log(0.000114394) = -39.416 dB

Second benefit of this change is a performance gain. I saw +0.15-0.3 Msamples/s in the benchmarks, except for the UC8 NODC that showed up to +2.8Msamples/s, on my Pi2 setup.

[mutability]

The two calculations are not equivalent, which is why you get different numbers. With a time-varying input it is not true that sum(|Xi|)^2 = sum(|Xi|^2)

[Mictronics]

Well, it depends what result is required. In case of mean power per sample block (mean envelope power) I think my method is just fine. For my understanding your method is calculating the mean of peak power of each IQ pair (mean instantaneous power).
Since mean power is going into the demodulator for noise calculation the question is if it's better to take the mean power of a whole sample block or the mean of peak per IQ. What is your experience on that?

[mutability]

The current implementation calculates power (in the only meaningful way for an unpredictably varying signal) and the demodulator expects to be able to deal with that as an energy value i.e. you can add and subtract in a simple way. Doing some type of averaging of the level (not power) over a sample block is not useful because the sample block boundaries are arbitrary and don't correspond with the received signal at all.

[mutability]

Or to put it another way: the power value reported for a sample block is the average power of that block, such that (power * duration of sample block) = total energy in the sample block. The demodulator measures the energy in the signals it finds and subtracts that from the total energy to find the remaining "noise" energy. Then turn that back into a power by dividing by the block duration.

AFAIK it is correct as it currently is and applying this PR would break it, so I'm going to go ahead and close it. If you do spot errors in the calculation, or a way to compute the same thing faster, let me know.