Compute SNR posteriors

[mpuerrer]

• Implement SNR posteriors for standard likelihood
• log_likelihood() now returns 3 floats (logL, rho_opt, rho_coh) and similarly log_likelihood_multi() will now return arrays of these quantities.

TODO:

• implement the SNRs for the marginalized likelihoods
• add tests

[stephengreen]

I have updated this PR with several changes:

• Made the code in core.likelihood more generic by not referring explicitly to GW quantities. The generic log_likelihood method now has an argument flag include_supplemental. If False (default), the method acts as before, returning a scalar log likelihood. If True, it returns a tuple containing the log likelihood, and the supplemental information as a dict.
• For GWs, the supplemental information includes the per-detector optimal and matched-filter SNRs.
• I modified the inner_product function to assume that the frequency axis is last rather than first. This allowed to streamline the code for the calibration-marginalized log likelihood, so I could unify it with the ordinary log likelihood.
• There are now unit tests comparing the log likelihoods from Bilby and Dingo, for IMRPhenomXPHM, SEOBNRv4PHM, and SEOBNRv5PHM models.
  • These create an injection at fixed parameters, and then calculate the log likelihoods at random parameter values. The threshold for comparison is 0.1, and this usually passes, but occasionally fails with slightly higher deviations. I'm not sure the cause of these small deviations.