Handle missing, NaN, Inf values in pd() and other summary functions

[bwiernik]

Currently, the posterior summary functions like pd() ignore missing and infinite values. These are retained (e.g., as missing) when compared against the null value, producing erroneous results like pd = 1 for a vector of all missing values.

bayestestR::pd(NA_real_)

x <- c(1, 2, NA, Inf, NaN, -Inf)
bayestestR::pd(x)

We should add an remove_na argument to these functions, defaulting to TRUE. If remove_na = FALSE and there are missing values, the functions should return NA_real_. When remove_na = TRUE, they should be removed before computing the result.

For Inf and -Inf values, arguably these are larger/smaller than 0, but they would usually mean some sort of convergence problem, so I'm thinking it might be better to remove them or to have an argument remove_inf defaulting to TRUE to control removing them.

[mattansb]

I think Inf/-Inf should remain - these can represent overflows (e.g., exp(big number)).

Implementation could be as easy as adding this to the underlying .p_direction() function, since all methods pass ... down to it.

Should the argument tbe named na_rm = TRUE for general consistency with R/tidy principals?

[strengejacke]

I think we have remove_na throughout our packages.

[strengejacke]

Maybe we can just give a message/warning for Inf values?