Fix bugs in _store_from_cache and repository.erase

[greschd]

The _store_from_cache needed to erase the content of its sandbox folder before copying over the content of the cache source. Currently, the existing contents are mixed with with the contents
to be copied in.
In fixing this, we discovered an issue in repository.erase, where an unstored node would try erasing the (inexistent) folder in the permanent repository, instead of the SandboxFolder.

Fixed in collaboration with @sphuber, so this should be reviewed by someone else.

[greschd]

I guess we should add some regression tests for this...

[greschd]

@sphuber do you also want a regression test for the Repository.erase part?

[sphuber]

@sphuber do you also want a regression test for the Repository.erase part?

Actually, a very simple one should not be too much work and probably give a lot of mileage. Just one where you store some files and call erase on unstored node and one where you call erase after storing (passing force=True). That should already cover quite a bit

[greschd]

@ltalirz in the test_store_from_cache (which is a new-style pytest test, yay) I now have to manually call reset_db, because some other part of the test suite (probably tests not using the fixtures) trashes the User.
Maybe it would make sense to also have a fixture which resets before the tests, not only after? It seems to be a valid pattern that you need the DB to be "pristine" for a certain test.

[greschd]

Docker check failing seems unrelated to this PR.

[greschd]

For the repository tests, I added 3 cases:

• erase unstored node
• erase stored node with force=True
• erase stored node (no force flag), check that it raises

[sphuber]

Thanks @greschd looks good, just request to add small comment in the code

[greschd]

Changed it to implicitly rely on #3783. Failing on purpose right now, but it should start working again once that is merged and this branch is updated.