Fixes/update for next project/known issues

[jessemorris]

• Projection covariance for point should be based on selected pixel covariance and depth covariance (as these are the actual measurements) and then popogated via the backProjection Jacobian. This will take a little bit of work to refactor
• As part of above refactor the feature API/frame API with so that all the information about projection/covariance depth etc can go straight into the feature. This include the way we currently update the depth of each feature from the frame. We should not need to do it like this - each feature should just have all the information in it.
• Python Bindings :)
• Update feature exactor matching to use SOA learned methods (e.g. super glue: https://github.com/yuefanhao/SuperPoint-SuperGlue-TensorRT/tree/main, or https://github.com/sair-lab/AirSLAM/blob/master/include/super_point.h). Try to use tensorRT in all these cases and make the GPU libraries optional dependancies. This is at least for static tracking - see the efficiency of dense tracking using these methods?