NavMap ObstacleFilter: optional obstacle memory

[fmrico]

Closes #125

With memory.enabled (off by default, so existing configurations behave as before), ObstacleFilter remembers what it has seen:

• Evidence in log-odds. Each NavCel accumulates evidence in obstacles_logodds, an int16 layer clamped as in OctoMap. A point above the surface is a hit. A cell is a miss when a ray from the robot to a point crosses it along the surface, or when a ground point lies on it.
• Decay. With memory.decay_time, cells that are no longer observed fade back to unknown.
• What the sensors know. obstacles_sensor holds obstacle, free or unknown for each cell.
• Composition. obstacles is the static map with the sensor layer on top (navmap::combine_layer, max). The sensors can mark cells that the static map does not know about, but they never erase a static obstacle. NO_INFORMATION is the declared unknown value of these layers.
• Layer identity. Both filters now detect a replaced layer through a weak_ptr instead of its address, so a cleared change record no longer causes a rebuild.

Cost per cycle, with a robot moving through home2 at 5 cm and a 720-beam laser: ObstacleFilter takes 1.07 ms with memory and 0.24 ms without.

Eight new tests cover:

• memory off by default;
• an obstacle out of sight is remembered;
• rays clear it within 2–3 cycles, and later cycles with the same evidence record no change;
• the sensors fill in unknown static cells and never erase a static obstacle;
• decay;
• clearing disabled;
• a new map resets the memory;
• invalid probabilities fall back to the defaults.

The simulation battery gives the same results in summit_warehouse_amcl, summit_bonxai_amcl and tiago_navmap_bonxai_amcl, both with and without memory (all reach the goal).