Minor updates

• Added instruction about how to train dm_control environments with minimal modifications.

• Remove AtariPreprocessing because it's merged as a PR in gym officially.

• Improve scripts & CI: use conda as much as possible for MKL optimization (like numpy/scipy) etc.