Count updates of DQN

[muupan]

This PR will add n_updates to the statistics of DQN. It is helpful for debugging.

I confirmed that

python examples/gym/train_dqn_gym.py --gpu -1 --replay-start-size 1000 --update-interval 1 --steps 2000

gets ('n_updates', 1000)

INFO:chainerrl.experiments.train_agent:outdir:results/20181024T204226.214715 step:2000 episode:9 R:-1.3913073996433227
INFO:chainerrl.experiments.train_agent:statistics:[('average_q', 0.3017159130785429), ('average_loss', 0.002146864709680402), ('n_updates', 1000)]

, and

python examples/gym/train_dqn_gym.py --gpu -1 --replay-start-size 1000 --update-interval 2 --steps 2000

gets ('n_updates', 500)

INFO:chainerrl.experiments.train_agent:outdir:results/20181024T204321.815223 step:2000 episode:9 R:-1.1019894359275235
INFO:chainerrl.experiments.train_agent:statistics:[('average_q', 0.25655420187569583), ('average_loss', 0.002052869248714256), ('n_updates', 500)]

.

[toslunar]

self.optimizer.t maintains this value and possibly more robust in serialization.

[muupan]

Good point. It uses self.optimizer.t now, and I confirmed that the printed statistics are the same as above.

[toslunar]

LGTM