Start model refactoring and fix tests

[swillner]

We started refactoring in models.py (more changes will follow). This is WIP, but as a first step to overall refactoring we think this could already be reviewed.

During the refactoring we found that in one place the calculation differed from those in other places.
In the original code for tasrange a factor 2 was missing when calculating the weights.

https://github.com/ISI-MIP/attrici/blob/8a735037f3d0e78f842e9646bcb66614179bc2df/attrici/models.py#L554C1-L554C90

The resulting changes in the output are minor, we adjusted the accuracy of the tests accordingly.

This PR also includes a fix for the tests to always run the tests (which are only run locally, not in CI, because they take some time to run) without using the cache.

[matthiasmengel]

I am not getting yet where the bug was.

            weights_nu_fc_intercept = pm.math.concatenate(
                [pm.Normal(f"weights_nu_fc_intercept_{i}", mu=0, sd=1 / (i + 1), shape=2)
                 for i in range(int(xf0v.dshape[1]) // 2)]
            )

seems to be consistent everywhere.

[swillner]

The difference is the missing 2 * in sd=1 / (i + 1) (vs. as in the paper sd=1 / (2 * i + 1)).

[SimonTreu]

For me the changes look good. The code structure is much clearer now and code duplication is avoided.