Support specifying fill behavior in frame_stacking

[ethanluoyc]

trafficstars

Add an option to allow for different frame_stacking behavior at the start of an episode.

Some original implementations of RL agents uses a different frame-stacking behavior at the start of the episode (repeat the first observation rather than zero pad). Add an option to support that. Keep the default to be the same zero-padding behavior to not break existing usage.