A3C LSTM GA for language grounding

[beduffy]

• Added A3C_LSTM_GA model from DeepRL-Grounding by Devendra Chaplot
• Contains cozmo_env PR
• Added checkpointing and experiment IDs of models
• Added tensorboardX which outputs to experiment folder with config+argparse settings
• Added avg_episode_return
• Added random actions on init
• Added 4 natural language tasks with variants specified by config files (NaturalLanguageLookAtObjectTask, NaturalLanguageNavigateToObjectTask, NaturalLanguagePickUpObjectTask, NaturalLanguagePickUpMultipleObjectTask)
• Added test case for 2 natural language tasks
• Added if statements to decide which model to use depending on whether task has natural language or not and therefore chooses A3C or A3C_LSTM_GA
• Added 3 config files and renamed main config to default_config.json
• Created task_utils.py for visualisation, preprocessing and reward functions
• Fixed config overwrite bug
• Fixed many other bugs I've forgotten about (e.g. one test case was broken)
• Renamed to PickUpTask from PickUp
• Added ability to remove lookupdown_actions and moveupdown_actions and put action alone (for pickup tasks but not put needed)
• Added ai2thor_examples.py for previous test cases that were purely just ai2thor run examples
• Added inbuilt_interactive_mode.py for script for exploring ai2thor with unity build path and .interact()
• Added pdb_interactive_and_check_bbox_reward_functions.py for testing reward functions and printing bboxes using pdb debugging
• VizDoom integration as originally with DeepRL-Grounding
• random scene reset choice and setting from config
• ASCII art figures for architectures within model.py
• Documentation