feat(models): add the StreamPETR detector

[vividf]

Summary

The StreamPETR detector: query-based head with a temporal memory bank, and the model wrapper with the deployment export ABI. Ported from the reference implementation (exiawsh/StreamPETR) with training parity as the goal.

Changes

• StreamPETRHead (models/detection3d/heads/streampetr.py):
  • Temporal memory: pre_update_memory aligns the bank into the current ego frame (timestamp/ego-pose bookkeeping), post_update_memory appends the top-k strongest detached proposals; prev_exists resets streams at scene boundaries; pseudo reference points seed cold starts.
  • Geometry-aware position embeddings (LID depth bins, frustum → lidar frame) with a float32 island — kilometer-scale translations exceed bf16 resolution.
  • Denoising queries with group attention masks extended over the propagated queries and temporal memory keys.
  • Hungarian-matched focal + L1 losses over every decoder layer; shared cls/reg branches across layers per the reference.
• StreamPETRDetectionModel (models/detection3d/streampetr.py): backbone/neck/head composition, grid-mask hook, memory reset on train/eval phase switches, backbone-vs-rest optimizer groups, and the three-module ONNX export (extract_img_feat / position_embedding / pts_head_memory) carrying the five memory tensors through the graph so the runtime manages the stream outside it.

Testing

tests/models/test_streampetr.py: forward/loss/predict shapes, memory recurrence across frames, prev_exists reset behavior, and export-wrapper vs eager-head consistency.

Stack 3/8 — base: feat/streampetr-model-components.