Learning a Prior for Monte Carlo Search by Replaying Solutions to Combinatorial Problems