Should provide additional conversions

[EricCousineau-TRI]

Follow-up to #153.

This is primarily in service of ROS 2 development in Anzu. Integration with Drake is a close second, though.
We should either provide these here, or as a package in this workspace.

We should provide both C++ and Python conversions, and exercise them.

• Python versions of geometry_conversions.h (either bindings, or reimplementations)
• Eigen matrices / NumPy arrays
• Point Clouds (\cc @hongkai-dai)
  • This could be Eigen matrices, and/or Open3D types.
• Images

\cc @IanTheEngineer @sloretz

[adityapande-1995]

Made an upstream ticket for ROS messages to be typecastable from python to C++ : ros2/rosidl#723

[adityapande-1995]

Update :

• Adding pybind11 bindings for geometry conversions #215 Adds bindings for geometry_conversions.h. Eigen matrices are bound to numpy matrices in simple cases, and as pydrake.common.eigen_geometry.Quaternion for cases where numpy does not have an equivalent data structure.
• [py] Add generic type casters #257 adds generic typecasters for ROS messages using serialization between rclpy and rclcpp.

[adityapande-1995]

@EricCousineau-TRI For pointclouds, IIUC, we need a conversion between the python and C++ ros message type (using something more efficient then serializing-deserializing) and then between the C++ pointcloud message and drake C++ type, right ?

If I look at geometry_conversions, what we are doing for say quaternion is :

• python ros msg -> C++ ros msg -> C++ drake type -> python drake type -> return to python world
• python drake type -> C++ drake type -> C++ ros msg -> python ros msg -> return to python world

Assuming we want the same for pointclouds, this would require 2 conversions :

1. C++ ROS PointCloud & PointCloud2 message <-> drake::perception::PointCloud
2. C++ ROS msg PointCloud & PointCloud2 <-> Python equivalents, not using serialization, but something more efficient.