Feature: do face recognition as a separate processor

[pozitronik]

Currently, each face processing model checks if the processed frame contains any faces. It requires loading a FaceAnalysis model and doing the same recognition for each processor in the chain.
It is possible to run face recognition once and skip all faceless frames later.

[pozitronik]

#133 also can be implemented here

[pozitronik]

I should discard the idea: the face needs to be recognized every time to process it (swap, enhance, etc.), so there can't be any optimization. However, it is possible to save recognition data from the first processor run and pass it through the processing chain.