feat: add MediaPipe face keypoint support

[JJassonn69]

This change add feature for adding face landmarks using mediapipe, the change matched the previous implementation for body and hands so the output can be skeleton pose compatible with openpose output to feed into the controlnet.

There was some performance degradation as reported by @ryanontheinside where the hands and pose detection was getting > 20 fps but with the addition of face landmarks the peformance is about 7-8 fps. We can do further testing to verify if the performance regression is the cause of main process of face landsmarks or the way the files are handled.

Also there is another route that can be followed where instead of monolith face, hand and face tracking, individual modules for each can be utilised such that if needed one can be activated and deactivated which might speedup the process according to the need.

This is a solution to the https://linear.app/livepeer/issue/ENG-2864/optimize ticket.