Fix various issues in TriggerRate

[rhaschke]

In order to actually reach the desired rate, the inequality needs to be replaced with greater equal.
Here is an example with publish_rate = 50 and Gazebo simulation (i.e. perfect sim time)

Before: rate = 47.619 Hz = 1/21ms

rostopic hz /franka_state_controller/joint_states
subscribed to [/franka_state_controller/joint_states]
WARNING: may be using simulated time
average rate: 47.619
        min: 0.019s max: 0.024s std dev: 0.00086s window: 143

After: rate = 50 Hz = 1/20ms (as expected)

average rate: 50.000
        min: 0.018s max: 0.022s std dev: 0.00063s window: 52

[rhaschke]

I found some more issues with this class:

• Using ros::Time() conversion to double can introduce rounding errors, causing sporadic misses in state publishing
• When ros::Time was reset (e.g. in simulation), there were no state updates at all, because now - time_stamp_ is negative.

[rickstaa]

@rhaschke Excellent observations! Much appreciated! I applied them to my fork.

@Maverobot Unfortunately, I couldn't test this PR on the real robot, but they look good. It would be great if these could be merged as they will significantly improve the performance ⚡.