Store information about which sampling method or algorithm was used

This was briefly discussed in #130 , but I am making it a separate issue so that the discussion does not get lost. I will copy what was discussed there:

It would be convenient to store information about which algorithm the elpd values come from. Since moment matching and refitting operate on individual observations, the information should be stored separately for each observation. For example, by default this value could be the same as is_method for all observations, but moment matching or refitting could change it from, say, psis to refit or psis-mm for some observations.

Details to discuss that come to my head at this moment

• Where to store it? pointwise is currently a matrix, could we make it a data frame to include a column with text? It could also be an element in diagnostics?
• What kind of information to include and how to name the different options (is method/moment matching/refit)? It seems that even is_method is currently not stored anywhere. Probably subsampling would be indicated here as well?

Start by reviewing the discussion in issue #130 and the current pointwise matrix and diagnostics structures mentioned here. Determine where per-observation sampling-method metadata should live and how methods such as psis, refit, psis-mm, and subsampling should be represented. Done means the storage location and naming scheme are agreed.