Cache improvements

[d70-t]

This PR

• uses the matching rules for restore_keys to be able to update the cache and thus fixes #104.
• It also adds caching for the conda environment (which uses mamba now) in a separate cache, thereby speeding up the environment creation.
• I thereby removed the python version from the build matrix, because mamba update seems to care only about what's written in environment.yml and we likely won't build the thing with different python versions concurrently.
• This change only upload the html part of the build output, which speeds up artifact creation and reduces required disk space. I don't think we ever used anything else, so it seems to be a more sensible choice.

In total, a build which fully utilizes the cache is now below 2 minutes or so and this number will often only increase by the amout required to execude the changed notebooks.

[d70-t]

As we ar currently struggling to run just one copy of the book, I guess testing multiple versions is still quite a bit away. If we'd want to combine testing multiple versions with caching properly, we likely would also have to pin all the dependency versions (currently it could happen that some cached notebook content has been executed using a different version of some libraries than newly executed content). I guess the current behaviour is quite ok though. At least we are checking that the notebooks work in principle. And if they break, the most likely reason is, that data moved away...

For the environment.yml, it currently says python>=3.9, which mamba interprets as "get the newest".

Thanks for reviewing! 😄