Fix locate_navcel on multi-surface maps and speed up queries

[fmrico]

Fixes #28

locate_navcel could return the wrong floor, and with a hint it was much slower than without one. Covering the fix with randomized tests turned up four more problems in the same code path; the list is in the issue comment.

Changes in navmap_core

• Closest surface in z. The vertical rays keep the closest hit over all surfaces, not the first surface in the list that is hit.
• Point exactly on a surface. The rays start 1 mm past p, so a point exactly on a surface (a robot at z=0, for example) finds that surface.
• O(1) owning surface. A NavCel -> surface index, built in rebuild_geometry_accels() and kept on copy, assign and move, replaces three linear std::find lookups. New public surface_of(cid) and kNoSurface. If the index is stale, it falls back to the linear search.
• Hint walk.
  • It steps across the correct edge (neighbor[(i+1)%3]).
  • It re-checks height_eps on the cell where it ends.
  • The hint and search paths both project the point vertically.
• Faster queries.
  • The closest-hit raycast visits surfaces in order of their AABB entry and stops once no surface can beat the best t.
  • The upward ray in locate is bounded by the downward hit.
  • BVH traversals use a fixed stack instead of a std::stack per query.
  • closest_navcel visits the nearer child first.

Tests

• test_navmap_properties: random multi-surface maps, with sloped and bumpy patches overlapping in XY and interleaved NavCel ids. It checks locate_navcel (with and without hints), raycast, raycast_many and closest_navcel against brute-force references, and checks adjacency symmetry. It also has robustness cases: NaN/inf queries, empty map and empty surfaces, degenerate triangles, points on vertices and edges, every cell as a hint, a cell in two surfaces, and repeated rebuilds.
• test_locate_multisurface: floors in normal and reversed order, points on, between, above and below floors, valid and invalid hints, hints on another floor, the hint walk, surface_of after copy, assign and move or with a stale index, and set_area / sample_layer_at on the right floor. The regression tests fail without the fix.
• Sanitizers: all navmap_core tests also pass under ASan/UBSan.
• Downstream packages: the tests of navmap_ros, navmap_examples and the EasyNav NavMap maps manager, planner and localizer pass unchanged.

Benchmark

Optional navmap_bench (-DNAVMAP_BUILD_BENCH=ON, off by default). 5 cm cells, before vs after:

Query	Before	After
locate with hint, 180k cells	2.8 us, grows with N	0.07 us, constant
closest_navcel, 180k / 540k cells	72 / 106 us	1.5 / 3.3 us
locate without hint, 3 floors, 540k cells	2.4 us, wrong floor	1.7 us

A deep copy costs about 10% more because it now copies the surface index.