feat: add ObjectMinPointsFilter transform to drop sparse GT boxes

[simonschlaepfer]

Summary

• Adds ObjectMinPointsFilter, a transform that drops ground-truth boxes enclosing fewer than min_points LiDAR points, filtering labels in sync. Motivation: boxes with too few points carry little geometric evidence, so training against them mostly injects noise. This is the analog of mmdetection3d's filter_by_min_points ground-truth-database option, exposed as a composable pipeline transform.
• min_points: int: a box survives iff its interior point count is >= min_points (strictly-less-than is dropped). 0 keeps every box; 1 keeps every box with at least one point.
• Counting is format-agnostic (incl. 9-DoF); a point in the overlap of several boxes is counted once for each box it lies in (membership, not first-box assignment), so a shared point can't push a box below threshold on a neighbour's account.
• Only the boxes/labels are filtered — the PointCloud3D and every other sample entry pass through unchanged. Follows RangeFilter3D's dict-based convention: accepts both a single sample dict and an (inputs, targets) pair.
• Absent points entry → no box has a defined count, so all boxes are kept; a present-but-empty point cloud counts as zero points per box (dropped unless min_points=0).
• Pairs naturally with ObjectPointsSample (feat: add ObjectPointsSample transform for per-object point thinning #133): after that transform simulates sparse returns by thinning each object to a target keep count, applying this filter with a matching min_points drops the objects sparse simulation pushed below the detectability floor. The two stay orthogonal — thresholds are matched by the user, nothing couples them automatically.
• Reuses the existing points_in_boxes_3d op (per-box counts are a sum over the [N, M] membership mask), so no new functional kernel was introduced.

Test plan

• pytest test/test_transforms_object_min_points_filter.py (40 tests: threshold boundary from both sides (min_points 0/1/2/99 over known per-box counts); overlap point counted for each box; box type/format preservation; format-agnostic across all four box formats; points and unrelated entries pass through unchanged; single-dict vs (inputs, targets) structure; absent-points keep-all vs empty-cloud drop-all; missing boxes/labels keys; empty boxes; min_points validation for negative/non-int/bool)
• ruff check and ruff format --check clean
• pyrefly check clean (0 errors)

[yeetypete]

Just some comments. Biggest issue is caused by me not using pytree in RangeFilter3D.