feat(BEVFusion): release BEVFusion 2.8.x

[KSeangTan]

Summary

This PR mainly releases BEVFusion/2.8.x with the following main changes:

• Update the training batch_size from 8 to 16. [1], [2]
• Add traffic_cone and barrier to output classes. [3], [4]
• Filter training bboxes by distance. [5], [6]
• Filter evaluation bboxes by number of points (>= 2). [6], [7]
• Add number of points per voxel from 10 to 32. [8], [9]
• Add projects/BEVFusion/bevfusion/bevfusion_voxel_encoder.py for custom voxel encoder layer, and simplify SinCos encoder. [10]
• Simplify sparse to dense in sparse conv encoder. [11]
• Purge custom LayerNorm in mmdeploy. [12]
• Simplify max-pooling and topk selection. [13]
• Add new TP errors to the documentations.

Performance

LargeBus

---

J6Gen2

---

JPNTaxi_Gen2

---

J6Gen2_Base

---

Base

---