Add ability to throttle covariance computation

[efernandez]

This PR adds the covariance_throttle_period ROS param to the fuse_models::Odometry2DPublisher to allow throttling the covariance computation. The motivation is to reduce the CPU usage by saving time not computing the covariance on every optimization cycle, i.e. every time notifyCallback() is called.

I've profiled the fixed-lag smoother, and the two major hotspots are Ceres solver optimization and the covariance computation, taking around half of the total computational time each.

The idea here is that we could reduce the CPU usage to almost half if we don't need the covariance at the same rate as the state. For example, if we set covariance_throttle_period to 2.0s, we get approx. half the CPU usage.

By default covariance_throttle_period is 0.0, which doesn't throttle the covariance computation, so we have the current behaviour.

Note that when the covariance isn't computed, we simply used the last one computed. We don't predict/propagate it forward (see code comment for more details and justification). It's still propagated if predict_to_current_time is enabled, but doesn't only for the small dt between the optimization and publish time.