`_save_metrics` assumes that `solve()` is called just once

[tomsilver]

but that's not the case when we're re-planning.

expensive way to reproduce:

python predicators/main.py --env sticky_table_tricky_floor --approach active_sampler_learning --experiment_id grid_row-planning_progress_explore --debug --strips_learner oracle --sampler_learner oracle --bilevel_plan_without_sim True --max_initial_demos 0 --sampler_mlp_classifier_max_itr 100000 --mlp_classifier_balance_data False --pytorch_train_print_every 10000 --active_sampler_learning_model myopic_classifier_mlp --active_sampler_learning_use_teacher False --online_nsrt_learning_requests_per_cycle 1 --max_num_steps_interaction_request 1000 --num_online_learning_cycles 10 --active_sampler_learning_explore_length_base 100000 --sesame_task_planner fdopt-costs --explorer active_sampler --active_sampler_explore_task_strategy planning_progress --seed 456 --execution_monitor expected_atoms

then look at the average number of predicates; it will be ridiculously high and non-constant, even though we're not predicate learning