Support for additional heads

[jaak-s]

I'd be interested to create environments that have add extra data for auxiliary training tasks available (e.g., reward prediction). Is that already possible, thx! I don't mind rewriting some methods to accomplish that if it is reasonably easy to do.

[takuseno]

@jaak-s Hi, thanks for the issue. Currently, the multi-head architecture is not planned yet. We could possibly do it by modifying algorithm classes.