profile and speed up ddpg with learned dynamics

[vlad17]

Here is the invocation of interest:

python cmpc/main_ddpg.py --model_horizon 10 --timesteps 15000 --verbose --seed 3 --exp_name timing --env_name hc --q_target_mixture --save_every 0 --explore_stddev 0.2 --mixture_estimator learned --ddpg_min_buf_size 10000 --dyn_min_buf_size 10000

We want to minimize the ddpg training time. For a first optimization pass, line_profiler-level profiling would be most appropriate. My guess is most of the time is being spent in the dynamics metrics, and there's lots of opportunities for speedups there since that class does a lot of costly numpy transfers.