DepthImageToPointCloud does not handle frame offsets

[RussTedrake]

@pangtao22 noticed that the RGB ad point clouds are off for the physical robot, and realized that our DepthImageToPointCloud class does not handle the case when the RGB camera frame is offset from the depth camera frame.

I'm not actually sure what the "right" algorithm is here (the projection seems nontrivial?). to leave myself a breadcrumb, the open3d version is here: http://www.open3d.org/docs/release/python_api/open3d.geometry.RGBDImage.html#open3d.geometry.RGBDImage.create_from_color_and_depth but opens with "RGBDImage is for a pair of registered color and depth images, viewed from the same view, of the same resolution. If you have other format, convert it first."

[EricCousineau-TRI]

From reviewing #14985, I believe the defect is as you mentioned; the depth image is not registered to same camera extrinsics nor intrinsics, due to this code snippet:

drake/examples/manipulation_station/manipulation_station.cc

Lines 167 to 186 in 8552564

	// To pose the two sensors relative to the camera body, we'll assume X_BC = I,
	// and select a representative value for X_CD drawn from calibration to define
	// X_BD.
	geometry::render::ColorRenderCamera color_camera{
	{renderer_name,
	{kWidth, kHeight, 616.285, 615.778, 405.418, 232.864} /* intrinsics */,
	{0.01, 3.0} /* clipping_range */,
	{} /* X_BC */},
	false};
	const RigidTransformd X_BD(
	RotationMatrix<double>(RollPitchYaw<double>(
	-0.19 * M_PI / 180, -0.016 * M_PI / 180, -0.03 * M_PI / 180)),
	Vector3d(0.015, -0.00019, -0.0001));
	geometry::render::DepthRenderCamera depth_camera{
	{renderer_name,
	{kWidth, kHeight, 645.138, 645.138, 420.789, 239.13} /* intrinsics */,
	{0.01, 3.0} /* clipping_range */,
	X_BD},
	{0.1, 2.0} /* depth_range */};
	return {color_camera, depth_camera};

Possible solutions:

• Register color to depth, e.g. via something like cv2.registerDepth. The math is not too hard to write in C++; dunno how great performance is.
• For the purpose of simulation, do not use baked-in different extrinsics and intriniscs; instead, make the same.

\cc @SeanCurtis-TRI