Change spin_thread default to false

[sloretz]

This changes the default value of spin_thread to false if the TransformListener is given a node. I'm assuming if a user has a node then they probably have an executor too. If so, then true for this parameter is potentially a problem because spinning on the same entities in two wait sets is explicitly undefined behavior in rcl

Noticed while looking at ros2/rcl#375.

[tfoote]

This likely works around the current issue. But it's not good practice for us to let people get into undefined behavior. In ROS1 this setup it's own callback queue. To that end we should do the same here. However I believe that is not available yet in the rclcpp API. It certainly wasn't when this was implemented.

The main reason for the extra thread being required is that the waitForTransform (now canTransform with timeout) methods would block any new incoming transforms thus guaranteeing that they would always timeout unless there's another thread processing the incoming transform subscriptions.

There's some todos re using the clock's sleep_for

geometry2/tf2_ros/src/buffer.cpp

Line 155 in 97a5741

// TODO(sloretz) sleep using clock_->sleep_for when implemented

but even more optimal would be something like a yield_for method that would put the thread back in the pool until the timeout instead of blocking/sleeping. However I don't know if that's technically possible.

In the mean time I think that a cleaner fix might be doing something such to making sure that the wait_list's don't overlap to avoid the undefined behaviors.

Edit: I found one releated thread: ros2/rclcpp#519

[sloretz]

I don't know what the right way to avoid spinning on the node in multiple wait sets is, but it sounds like this PR isn't it.