Adding ML-based occupancy prediction to Autoware

[lexavtanke]

As mentioned in the #3607 I've proposed to add ml based occupancy prediction to Autoware stack as this is more modern method and it can enhance capability of the system.

As now this is pretty modern method I offer to integrate it as optional component to perception part. Here is new architecture

So users will be able to choose which occupancy grid map they want to use and we can easily compare the results.
Module will consume lidar and/or camera data (because most of the modern methods are camera based) and output occupancy grid map.

This loses semantic information but allows more easy integration and provide ability to test in real conditions.

Further with good results in grid map we can incorporate occupancy prediction as alternative or supplement to clustering for the tracking.
Also further can be discussed integration semantic information to grid map or use 3d volumetric data in planning pipeline.