Add DALOcc occupancy-aware head and detector with config

[thelastsun120]

Motivation

• Introduce a sparse pillar occupancy branch for DAL to provide occupancy supervision that can improve classification/heatmap cues while keeping box regression LiDAR-only.
• Expose occupancy-specific training wiring so voxel semantics and camera masks can be passed through the detector into the head.

Description

• Add DALOccHead in mmdet3d/models/dense_heads/dal_occ_head.py that extends DALHead and implements occupancy proposal, sparse pillar decoding, feedback mechanisms (occ_feedback modes), and occupancy losses (loss_occ_proposal and loss_occ).
• Add DALOcc detector in mmdet3d/models/detectors/dal_occ.py which wraps DAL and forwards voxel_semantics and mask_camera into the head loss computation.
• Register new modules by exporting DALOccHead in mmdet3d/models/dense_heads/__init__.py and DALOcc in mmdet3d/models/detectors/__init__.py and add a base config configs/dal/dal-occ-base.py with default occupancy hyperparameters.
• Implement occupancy target building, pooling of camera masks to occupancy bins, sparse pillar selection/decoding, and wiring of occupancy losses into the head loss method.

Testing

• Performed a basic import smoke test with python -c "from mmdet3d.models import DALOcc, DALOccHead" to verify registration and imports succeeded.

---

Codex Task