BUG: Transfer of covariate embedding is broken

[siboehm]

We need to invest more effort to get the covariates embeddings to properly transfer from LINCS ⇒ SciPlex. Picture to illustrate:

The solution is to save the string name of the covariate with the embedding into the state_dict, st we can keep the same mapping. The way we currently have it implemented we just pull a basically random embedding from the top of the table, so this is not super useful.

Needs to support:

• Extension of covariates (with a warning): Adding more rows for covariates that haven't been seen before. Maybe init to the mean of the others?
• Contraction of covariates: Returning the same covariates (same ordering)

[MxMstrmn]

Atm, we do not have the covariate embeddings available for the pretrained LINCS models, do we?

But I agree, this is missing in the transfer at the moment.