ensure likelihood convergence

This MR adds a new likelihood class that enforces a minimum number of effective samples per event to ensure good convergence of the Monte Carlo integrals. This is entirely analogous to the method already widely applied to the sensitivity Monte Carlo integral.

We require that all of the single-event marginalized likelihoods have n effective > n events.

This is a comparison of the difference this makes (event lists may not be exactly the same, but are close).

Without the convergence check, we see that a very high likelihood mode is found with small values of the various sigma parameters. This mode is completely absent when requiring convergence. I recommend using this method to remove the spurious mode.

Without convergence check:

With convergence check:

It does non-trivially change the prior distribution (in a data-dependent way!), but we can evaluate this numerically and for most parameters, there isn't a visual difference. I've pasted in prior comparisons for one set of input events below. The specific parts of the space cut will depend on the specific events we are considering.

(The change in mmin is due to this test necessarily failing where there are no samples with non-zero support for at least one event, this region has zero likelihood and so isn't impacted by this.)