Perform likelihood calculations on log scale to improve accuracy in edge cases

[AngusMcLure]

The fi_pool_cluster() currently does a lot of integration over integrands involving lots of subtraction, powers, and multiplications, which begin to suffer when prevalence (the variable of integration) is close to the 0 or 1. Changing these calculations to be robust at these edge cases might avoid some of the convergence issues we are getting with extreme values of beta. I have already done this for logitnorm and this looks promising, but should apply this more broadly. A list of functions consider this for:

• FI Integrand functions (link-norm)
• phi and 1-phi
• FI Integrand functions (beta)
• Change of variable integrand functions (link-norm only)
• link functions
• inverse link functions?
• derivative of inverse link functions?