NavMap AMCL: per-particle heading noise; IMU turns, not its absolute yaw

[fmrico]

Closes #131

• Noise. Each particle gets its own heading noise in the prediction.
• IMU. With an IMU, the particles take its roll and pitch and turn by the change of its yaw (a gyro turns far better than skid-steer odometry). They keep their own heading; the IMU's absolute yaw is never used.
• Without an IMU, the orientation follows the NavMap surface and the odometry's turns, as before.

In the Summit excavation, on a 50 m route with an IMU and slopes of up to 22°, the route completes in all runs. The mean error is 0.41–0.49 m and the maximum 0.87 m, against 0.9–1.23 m before, and the result no longer depends on Gazebo's exact IMU yaw. The wheel odometry drifts 30 m over the same route.

Three new tests cover:

• each particle spreads in heading while turning;
• the IMU's roll and pitch are used, its yaw changes are followed, and its absolute yaw offset and the ±π wrap are ignored;
• moving on a tilted robot stays on the surface.

The first two fail on the previous code.