Improved 2 metrics, removed 1, added 1

[AntoinePassemiers]

"recovery_2" metric:

• Found 3 critical bugs, so I reimplemented it from scratch.
• Based on a consistency score, defined based on the spread between maximum and minimum Spearman correlation coefficients across groups (donors).

"anchor_regression" metric:

• Removed it.

"regression_3" metric:

• Added a new "regression_3" metric to replace "anchor_regression". It is based on residual correlations. Core idea, in short: after fitting gene expression on the expression of selected TFs, the less the residuals are correlated with the non-selected TFs, the better.

"vc" metric:

• Fixed the same issue as in the "sem" metric: the inferred GRN was accidentally shuffled
• Predict perturbations (x_pert - x_control), rather than the expression of perturbated samples (x_pert)
• Initialize the NN and baseline NN with the exact same parameters (except the GRN weights)
• Use perturbation embeddings as linear operators to be applied in the latent space
• Fixed mathematically wrong implementation choices in GRNLayer
• Use stratified group K-fold for CV, where groups are (perturbation, cell type) pairs
• Improved the NN architecture and added strong regularization
• Keep track of the best solution
• IMPORTANT NOTE: I tested the metric on the small OP dataset, predictive performance was really low and final scores very noisy. I attribute this to the small number of samples, and hope performance will scale with the dataset size.