Migrate Cineast's VisualTextCoEmbedding

[ppanopticon]

Task Description

In #94 we discussed features that should be ported over from Cineast. One of these features was the VisualTextCoEmbedding feature that can be used to query image and video content using textual queries.

While the current implementation is arguably worse than OpenCLIP, is has one advantage: I uses information not just from the keyframe but also surrounding frames and may therefore spot things, other models might miss.

The goal of this issue is therefore to migrate the VisualTextCoEmbedding feature so that it can be used until we have a better alternative.

Dependencies

In light of #19, we ought to check, whether this should be part of the FES infrastructure or TorchServe.

Boundary Conditions

None