Large memory use and very slow post-processing with pool

I noticed that the parameter reconstruction was taking a huge amount of time and memory for some BNS analyses.

Other users have mentioned similar issues.

My suspicion is that this is because of overheads in transferring the likelihood object through the multiprocessing pool repeatedly (e.g., here).

We can get around this with the same method used for sampling. I suggest waiting after !1043 (merged) is merged so we can use the _sampler_convenience_dump.