Use hook functions

[muupan]

This PR add StepHook interface to customize training loops, e.g. decaying or increasing some parameters. Learning rate decaying in chainerrl.experiments.train_agent_async is now implemented using a hook.

Resolves #84

[coveralls]

Coverage increased (+0.1%) to 71.175% when pulling 0219fe5 on muupan:hook into bf46346 on pfnet:master.

[coveralls]

Coverage increased (+0.1%) to 71.188% when pulling 16f3494 on muupan:hook into bf46346 on pfnet:master.

[coveralls]

Coverage increased (+0.1%) to 71.188% when pulling 5d7e7cc on muupan:hook into bf46346 on pfnet:master.