Overhaul of the Python API

[nagitsu]

Adds several utility functions to ease the use of Luminoth through Python:

• Detector (in luminoth.tasks) to abstract the usage of object detection.
• vis_objects to visualize the results by drawing the bounding boxes in the image.
• read_image to read images from disk without needing to know about Pillow.

With this, the following is now possible:

from luminoth import Detector, read_image, vis_objects

image = read_image('cat.png')

# Assumes `accurate` checkpoint by default. Other checkpoint, or a config object,
# can be specified.           
detector = Detector()
objects = detector.predict(image)

print(objects)

vis_objects(image, objects).save('cat-out.png')

Further work is required along the following lines:

• I/O functions to read batchs of frames out of a video.
• Remove the TODOs in Detector that depend on rewriting PredictorNetwork.
• Provide a way to access the detected objects' embeddings for further processing.
• Document the new API.
• Add examples on processing both images and videos.