[ilia-kats]

Same as scverse/anndata#2707, but for MuData.

There are several hacks to bypass the sanity checks in anndata's implementation:

• overriding the attrs with something returning anndata's encoding type and version. This is necessary to be able to pass in a MuData object to anndata's write_use, which happens when writing e.g. MuData.obs, MuData.var etc.
• Creating AnnData accessors from MuData accessors. This is necessary to bypass the dimensionality check: The check reads obs_names / var_names of the passed in object, which would be a MuData object if using a MuData accessor, so the names don't match what's on disk for the individual modality.

Removing the hacks would require making anndata's write_use more flexible: Adding a boolean argument to disable encoding type/version checking and adding arguments to pass in obs_names and var_names to check against.

The CI is failing due to warnings related to Pandas 3.1. The warnings will disappear with anndata 0.13.5. Everything else is passing (though the new code is not tested in CI yet becase there is no pre-release of anndata 0.14, but the tests pass locally).