goal conditioned GoalTask and a few other small improvements

[le-horizon]

the few improvements including

1. improving camera position to avoid penetration into other objects
2. allow non-ego centric full states observation in GoalTask
3. log info observation_desc

[le-horizon]

where is observation_desc logged?

Currently I'm logging.info the obs desc.

[le-horizon]

Thanks for the careful review @Jialn . Done or answered all.

[le-horizon]

Merged in the changes. I also added 1) multi dim reward 2) auxiliary achieved field 3) made human rendering consistent with rgb_array mode (in terms of pseudo randomness)

Please take another look.

Thanks, Le