Add minimal novel package discovery

[bollwyvl]

Problem

At present, it is hard to calculate the exact package versions and URLs needed to build a reproducible site.

Proposed Solution

Given a kernel session has installed a couple packages, offer something like piplite.freeze which would:

• call micripip.freeze
• weed out all of the packages that weren't actually installed (not everyone needs galpy and scipy, but some will)
• inject the packages needed to get up to pyodide-kernel
• emit a repodata.json that can be dropped into {lite_dir}/pypi/repodata.json
• enjoy getting import foo without having to %pip install foo

Additional context

• extracted from Generate pyodide repodata, add more mocks #27 (comment)

A much-heavier-duty version of this might try to use micropip, depfinder and/or conda-lock (which bundles poetry) to do this in a browserless, nodejs-less command line setting, but interactive-first is never the wrongest way to go.

[rth]

Yes, I confirm this functionality would be very useful to smooth the difference between notebooks with in-brower and out of browser kernels.

For instance, with the deployment of Jupyterlite in the scikit-learn example gallery (dev deployment only for now), it's a bit confusing to users why some packages can be just imported and others need to be pip installed first.

[rth]

inject the packages needed to get up to pyodide-kernel

Do you mean copy the packages from wherever they are into the deployed site? For reproducibility yes, that would be good.

However, there is also a use case in the other direction where you want to add more PyPI packages to the pyodide-lock.json (pointing to the remote URLs where they are) without actually copying those files locally. For instance, scikit-learn deploys docs into a Github repo (that gets then deployed via Github Pages) and they don't realistically want to push 10+MB of packages there.