feat(streampetr): add auxiliary 2D head, partial-ignore, and DDP loss normalization

[vividf]

Summary

The training-quality core of the stack: auxiliary 2D supervision, partial-ignore for sparsely annotated classes, and DDP-correct loss normalization.

Changes

• FocalHead2D (models/detection3d/heads/focal2d.py): Focal-PETR-style dense head over the stride-16 neck features — per-token class scores, centerness, LTRB boxes, and projected 3D centers, supervised by quality focal, weighted L1, GIoU, center L1, and Gaussian-heatmap centerness losses. Cost/loss weights follow the reference recipe (2/5/2/10 + 1); the default HungarianAssigner2D mirrors the same cost weights. Loss-only: inference and deployment never run it. The reference's top-k token pruning is intentionally not ported.
• CPFPN (models/common/necks/cp_fpn.py): the reference StreamPETR neck (1×1 laterals, nearest top-down, 3×3 refine on level 0), with mm-compatible parameter names.
• Partial-ignore (models/detection3d/partial_ignore.py): traffic_cone/barrier handled purely as classification-column weights — on an un-annotated frame, every query/token drops only the ignored class columns. Structurally unable to perturb Hungarian matching, bbox targets, or the normalization factor. Both heads share one masking helper and the same semantics; missing or mismatched status flags raise instead of silently training the ignored classes as background.
• DDP loss normalization (task_modules/streaming.py + both heads): positive counts are averaged across ranks (mmdetection's reduce_mean), so the sharded loss equals the single-batch equivalent regardless of GPU count. Every collective sits on a rank-uniform code path — a rank whose batch has no ground truth still participates in the DN count collective with a zero contribution, so mixed-GT steps cannot deadlock.

Testing

Real multi-process gloo tests (test_streampetr_loss_distributed.py): a sharding-invariance property test (DDP-averaged sharded loss == combined-batch loss), the mixed-GT deadlock regression, and reduce-mean semantics. Partial-ignore has unit tests down to exact masked rows/columns and an end-to-end head.loss wiring test.

Stack 7/8 — base: feat/streampetr-2d-annotations.