* Support asynchronous unroll

Perform unroll and training concurrently. This can be useful in two
situations:
1. The environment step is expensive. This can help reducing time
spending on unroll
2. Realtime training where the interaction with environment happens
in realtime. For this, we can set TrainerConfig.unroll_step_interval
to the desired interaction period.

* Fix circular import

* Address comments

* Fix failure for build.sh

Need to avoid mess up with commandline arguments

* Address comments