[models] Add keypoint detection models

[frgfm]

Currently we perform keypoint detection by:

• using a robust image classifier
• use its class activation to extrapolate the keypoint

We should explore alternative ways with light backbones to perform actual keypoint detection and train it properly.

[Aym98]

Is this issue still relevant ?

[frgfm]

Hey @Aym98 👋

Yes, so far we've defaulted to object detection, but this is still not ideal. Add comparatively, keypoint detection is a much more adequate task to model. We could train much lighter models for this 👌