[integration/peerlab] Bugfix/tiling issue

[Tobiaspk]

Closes #40. This PR makes building QuadTree constructino more robust. This issue is common across datasets, but particularly often seen when processing the Atera Breast Cancer public dataset.

This PR fixes the error where -1 is passed into bincount.

Cause: cuspatial sometimes generates invalid QuadTrees where tiles overlap across levels, while other regions are not covered at all. This results in invalid tile assignments during spatial queries. This issue appears randomly despite the pipeline being deterministic, and occurs in roughly 2–35% of Segger runs. In some datasets, such as Atera, even 100%.

Fix:

• Add is_quadtree_valid: Simple validation that all points are covered by the generated quadtree.
• Add margin_bounds to get_quadtree_kwargs: Expands the bounds used to construct the quadtree. This produces more valid trees.
• Retry quadtree construction with larger max_size: If a generated tree is invalid, rebuild with an increased tree size. Particularly important for _points_in_polygons_contains in query.py, which defaults to max_size=10000. In the Atera dataset this frequently produces invalid trees.

Example log of dynamically increasing tree size:

2026-05-17 23:08:45 | DEBUG    | segger.geometry.quadtree:178 | GPU: 20.61 GB (peak 67.49 GB) - Building quadtree on 624327019 points with max_size=10000, max_depth=14
2026-05-17 23:08:46 | WARNING  | segger.geometry.quadtree:196 | GPU: 23.11 GB (peak 67.49 GB) - Invalid quadtree generated with max_size=10000. Retry with max_size=20000.
2026-05-17 23:08:46 | DEBUG    | segger.geometry.quadtree:204 | GPU: 23.11 GB (peak 67.49 GB) - Quadtree built: 75131 leaves