Cache loaded source data to reduce memory footprint

[JelleAalbers]

This lets Source keep a dictionary cache of loaded source-data (like PDFs), which reduces memory footprint if the same source is being loaded many times (for example, in models with many nuisance parameters that don't affect all sources). Thanks to @kdund for spotting and correctly diagnosing the problem.

I also included two bonus features I've been meaning to commit for a while:

• Skip pdf_task file creation if you're doing single-core computation. You can just as well compute the PDF immediately.
• best_anchor inference method, which finds the shape parameter anchor combination that best matches the data. Useful as a guess for further fitting.

[kdund]

Checked with 3d likelihood using binned histograms-- ran as expected, and decreases memory use.

Best_anchor checked for run, sensible returns, but no detailed checks performed.

Caching in the source avoids two problems my solution had:

• I wrote a separate function (via copy-paste) to duplicate the hash, which turned out to not work
• it will cache pdfs used by multiple likelihood once