Debiased maximum-likelihood estimators for hazard ratios under machine-learning adjustment