feat: collision check for data augmentation

[rej55]

Description

Overview

Adds a validity filter to StatePerturbation (ego-state perturbation for data augmentation) that rejects perturbed samples where the augmented ego position would cause a physical collision.
Without this filter, a randomly perturbed ego can overlap with surrounding objects or cross a lane boundary, introducing physically impossible training examples.

Changes

diffusion_planner/utils/data_augmentation.py

New module-level geometry utilities (pure functions, no side effects):

Function	Role
_cross2d	2D cross product — core primitive for intersection tests
_rect_corners	Converts (x, y, cos_h, sin_h, length, width) → 4 corner points [B, 4, 2]
_sat_signed_distance	SAT (Separating Axis Theorem) signed distance between two rectangles; negative = overlap
_segments_intersect_rect	Checks whether any of N line segments intersects a rectangle, using both edge-crossing and point-in-polygon tests

New StatePerturbation._check_aug_validity method:

Returns a [B] bool tensor (True = invalid) based on two conditions:

1. Ego vs. neighbor polygon — SAT signed distance < 0 for any valid neighbor agent (uses neighbor_agents_past[:, :, -1, :] for the current timestep; per-agent size from feature indices 6/7)
2. Ego vs. lane boundary segments — left boundary (lanes[..., :2] + lanes[..., 4:6]) and right boundary (lanes[..., :2] + lanes[..., 6:8]) consecutive waypoint segments that intersect the ego rectangle. Boundary sides with offset norm ≤ 0.01 m are treated as absent and skipped.

Ego vehicle size is read from inputs["ego_shape"][:, 1:2] (length) and [:, 2:3] (width) when available, with fallback to module-level constants EGO_LENGTH = 5.0 and EGO_WIDTH = 2.0.

StatePerturbation.augment update:

collision = self._check_aug_validity(ego_current_state, inputs)
aug_flag = aug_flag & ~collision

Samples flagged as invalid retain their original (unperturbed) ego state.

diffusion_planner/tests/test_data_augmentation.py

Added 33 new tests (26 → 59 total), covering:

• _cross2d, _rect_corners, _sat_signed_distance, _segments_intersect_rect
  • unit tests for all geometry primitives
• _check_aug_validity
  • neighbour overlap/clear, left/right boundary cross/clear, zero-offset ignored, ego_shape override, batch mixed
• augment integration
  • collision suppresses aug_flag, no-collision preserves it, batch selective suppression

How to test

Run the unit test suite:

cd diffusion_planner
python tests/test_data_augmentation.py
# Expected: Results: 59/59 passed, 0 failed

# or with pytest
pytest tests/test_data_augmentation.py -v

[SakodaShintaro]

I've started the training. I'll check the results and merge it if there's no deterioration.