[fmrico]

Closes #165.

Rewrites easynav_mpc_controller as a linear time-varying MPC solved as a sparse QP with OSQP (osqp_vendor), in place of NLopt.

• Motion models: diff, omni, ackermann and aerial (x, y, z, yaw), linearized around a nominal trajectory: the previous solution shifted, or the reference when the robot lags it. A time_constant models actuator lag.

• Constraints:

  • velocity and acceleration limits from robot_limits, including the vertical ones;
  • the turning radius (|w| ≤ v / R);
  • one half-plane per nearby obstacle and step, from the footprint plus safety_margin.

  Bounds always include the nominal, so the QP stays feasible.

• Reference:

  • the path is projected forward only;
  • its speed profile is limited by curvature, lateral acceleration, the goal and obstacle stops;
  • for turning-radius robots, the path is smoothed adaptively.

• Topics: <node>/<plugin>/predicted, /reference and /obstacles.

• Parameters: renamed parameters still work with a deprecation warning. safety_radius, use_collision_checker and verbose are no longer used and only warn.

• OSQP versions: works with OSQP 0.6 (Humble, Jazzy, Kilted) and 1.x (Lyrical, Rolling).

It needs the new robot_limits fields from #156 (max_vertical_vel/acc/decel, min_turning_radius), so CI will fail until that is merged.

Tests: unit tests for the model, QP and reference; 13 closed-loop scenarios with a simulated robot (corners, narrow passages, obstacles, Ackermann turns, aerial climbs); and parameter validation.

In Gazebo:

• Kobuki: 8/8 goals.
• Omni: 6/6 goals.
• Prius: a lap of Sonoma at ~7.5 m/s, with a mean lateral error of 0.08 m.
• UFO: across a town, 10/10 waypoints.

Builds: Humble, Jazzy, Kilted, Lyrical and Rolling.