Draft: Fix reduced chisq calculation

Related to #49

See the issue for a full description.

As with !100: this MR is mostly to share the changes for testing and further development. Don't review the changes until we've had a discussion/test of the science results.

In short, where previously we calculated cmbchisq as: chisq_A/autocorr_norm_A + chisq_B/autocorr_norm_B + chisq_C/autocorr_norm_C in postcoh_kernel.cu and THEN divided it by nifo in postcoh.c

we now instead calculate it as: (chisq_A + chisq_B + chisq_C) / (autocorr_norm_A + autocorr_norm_B + autocorr_norm_C)

The only way the old calculation makes sense to me is if autocorr_norm_A, autocorr_norm_B, and autocorr_norm_C were all 1 before. We'll need to check where we sourced our autocorr_norm matrix from before...