Terrain LOD refactor: edge stitching, hybrid mesh LOD, performance fixes

[brendancol]

Summary

• Edge stitching replaces tile skirts: boundary vertex Z values are interpolated from the coarser neighbor's pyramid data via _stitch_tile_boundary(), eliminating T-junction cracks without extra geometry. _needs_stitch() fast-check avoids vertex copies when no neighbor differs.
• Hybrid LOD for placed geometry: _MeshChunkManager applies simplify_mesh() during merge based on chunk LOD level (LOD 0 = full detail, LOD 1 = 50%, LOD 2 = 25%, LOD 3+ = 10%). Lazy cached in _simplify_cache, evicted when chunks leave visible set. Curves excluded.
• Critical bugfix: _rebuild_vertical_exaggeration() was missing cache_key definition — caused NameError crash on VE change (Shift+Home/End) with non-LOD terrain.
• Performance fixes: early frustum culling before LOD/roughness computation (~40-60% fewer evaluations), single-pass future harvesting, O(1) stale tile eviction, simplified Z re-snapping (final_z = stored_z * ve), GPU threshold bumped 1K→10K, in-place cache eviction, pyramid level eviction above max_lod, float32 roughness computation.
• Robustness: build retry budget (max 3 attempts with logging), I/O prefetch rate limit (max 8/tick), get_metrics() structured API, _has_in_flight_work rename for clarity.
• Code cleanup: removed dead _add_tile_skirt function, extracted update() into focused helper phases (A-E), updated docstrings and mesh.py references.
• Tests: 114 tests (up from ~30) covering edge stitching Z-value verification, hysteresis dead zone, retry budget, mesh simplification, heightfield integration, streaming, batched upload, and roughness-adaptive LOD.

Test plan

• pytest rtxpy/tests/test_lod.py — 114 tests pass
• Manual: toggle LOD (Shift+A) on a large zarr terrain, verify no T-junction cracks at tile boundaries
• Manual: change VE (Shift+Home/End) on non-LOD terrain — should no longer crash
• Manual: verify placed geometry (buildings/roads) simplifies at distance with chunk manager active
• Manual: streaming mode — verify I/O prefetch doesn't flood thread pool on initial load