Use same actor kind when BC and training

[mtrofin]

The BC policy is trained using a QNetwork actor, which is different from the one we use for normal training. See for example:

ml-compiler-opt/compiler_opt/rl/inlining/gin_configs/behavioral_cloning_nn_agent.gin

Line 19 in 9d00bcf

create_agent.policy_network = @q_network.QNetwork

ml-compiler-opt/compiler_opt/rl/inlining/gin_configs/ppo_nn_agent.gin

Line 35 in 9d00bcf

create_agent.policy_network = @actor_distribution_network.ActorDistributionNetwork