On-Device pointcloud generation from StereoDepth node

[asa]

Currently the device has a calibration matrix and outputs depth data which must be fused on the host into a pointcloud.

It would be advantageous for various on device pipelines to be able to calculate a 3d pointcloud using these two components directly using shaves at high speed. Currently there is a proof of concept for using pytorch to create a pointcloud , which achieves 26fps at 400_P.

It would be ideal to enable an output for the StereoDepth node which would directly compute this output in an efficient way using shaves. It boils down to a simple matrix multiplication.