`predict` on new samples using threshold policy

[erikcs]

Currently, predict returns the underlying allocation for the policy estimated (and evaluated) on the test set. Another use case arises when we have new CATEs and cost and need to predict treatment allocations for these samples under a given budget constraint. In this case, a fitted maq object should use the estimated slope $\rho$ cutoff value $\lambda$ from the test set to determine and return the appropriate allocation (e.g., use the thresholding policy in Theorem 1 here https://arxiv.org/abs/2306.11979)

Doing this requires storing the $\lambda$'s as part of the C++ solver's solution path.