(suggestion from the opam-repository bot)

Without the hack, the test run appears to cache properly,
and some parts of the build are cached as well (but not all),
and building the documentation does not get cached.

This hack does appear to make a difference, in that it ensures
that the documentation gets cached properly. The build still
suffers from imperfect caching.

The problem with the caching key is that it hardcoded details about
our project (path_glob.opam); people would need to change the name to
adpat it to their own project, making the CI script less portable.  In
addition, Gitlab advertises limitations on how many files can be using
as caching keys (at most 2??), so this would be an issue for people
willing to list several .opam files.

Gathered from bits and pieces at
  https://discuss.ocaml.org/t/basic-gitlab-ci-configuration/3327
and lots of CI golfing.

(Report from hackwaly)