[FEATURE] Make `balance` API closer to the sklearn syntax

[talgalili]

The current workflow of just calling adjust might not be the easiest API for general Python/sklearn users.

[talgalili]

The main API to consider are: transform, fit, predict.
Several design options should be considered.
E.g., we want transform applied for both sample and target. We want fit based on sample for outcome model and sample+target for survey weights (IPW). Then we want predict (predict_weights?) for sample, but predict for target for outcome modeling. And also predict for both sample+target as a pre-step for getting proxy features for use in NMAR.

[talgalili]

AFTER the current issue is resolved, I think it would be easier to work on:
#155

[cneuralnetwork]

@talgalili added a PR to solve this!