Fix reduced cmbchisq calculation

In postcoh_kernel.cu, we currently calculate the reduced cmbchisq to be the sum of the reduced chisq of each detector.
However, the correct calculation is to take the sum of the chisq of each detector, and reduce that (using the sum of their autocorr_norm values).

See spiir-production-Oct27-dev (moved to tdavies__fix_reduced_chisq_calculation)

This was implemented and tested in O3, and found to be of low significance. It's not worth doing a week long injection test on this change alone, but it could be included with others.