Add FaceDetector class

[DavidLandup0]

Requirements:

• CNN-based (not feature + SVM based, like Dlib's original frontal detector)
• CUDA-enabled
• Both TF and PT backends
• Real-time (i.e. at least 30FPS on consumer-grade GPUs)
• Allow inference on batched input

Potential architecture: YuNet (like in Kornia), though, could we use different architectures and/or backbones to be specified by the user?

API idea:

detector = deepvision.models.FaceDetector(nms_threshold=0.1,
                                 confidence_threshold=0.5,
                                 limit=5,
                                 architecture='arch1', 
                                 backend='backend')
dets = detector(img) # img.shape = (B, C, H, W) or (B, H, W, C)