Fix provenance in recipe_multimodel_products

[jvegreg]

Another very complex fix

See #1713

Tasks

• Create an issue to discuss what you are going to do, if you haven't done so already (and add the link at the bottom)
• Give this pull request a descriptive title that can be used as a one line summary in a changelog
• Make sure your code is composed of functions of no more than 50 lines and uses meaningful names for variables
• Circle/CI tests pass. Status can be seen below your pull request. If the tests are failing, click the link to find out why.
• Preferably Codacy code quality checks pass, however a few remaining hard to solve Codacy issues are still acceptable. Status can be seen below your pull request. If there is an error, click the link to find out why. If you suspect Codacy may be wrong, please ask by commenting.
• Please use yamllint to check that your YAML files do not contain mistakes
• (Only if really necessary) Add any additional dependencies needed for the diagnostic script to setup.py, esmvaltool/install/R/r_requirements.txt or esmvaltool/install/Julia/Project.toml (depending on the language of your script) and also to package/meta.yaml for conda dependencies (includes Python and others, but not R/Julia)
• If new dependencies are introduced, check that the license is compatible with Apache2.0

Modified recipe/diagnostic

• Update documentation for the recipe to the doc/sphinx/source/recipes folder
• Update provenance information if needed
• Assign the author(s) of the affected recipe(s) as reviewer(s)

---

Closes #1713