CI: more robust GitHub actions caching

[alexsavulescu]

Currently we only rely on a simple keyword + date for caching.
The idea is to somehow make use of ModelDB API so that we know when to retrieve (updated or new) models? Otherwise add a manual trigger?

[olupton]

Based on https://neuron.yale.edu/ftp/neuron/2019northwestern/python-2019-10-18.html it may be possible to use queries like https://senselab.med.yale.edu/_site/webapi/object.json/12631?woatts=23,24,25,26,27,28,112,113,114,155,249,299,471,488,502,582 to get the last-modified date of a model without downloading its content. It's not immediately clear how to query many models at once.

[alexsavulescu]

Ideally we'd get a unique API call with latest update times (like we have for getting all NEURON accession numbers https://github.com/neuronsimulator/nrn-modeldb-ci/blob/master/modeldb/config.py#L6 ) but I remember I couldn't find any.

[olupton]

This was addressed by #88.
Now in each run we make two API calls, one for the full list of models and the second for the full list of their modification times.
The cache metadata also includes the modification time, and this is used to only re-download new or updated models.