fix(t4metric_v2): filter GT by annotation num_lidar_pts

[vividf]

IMPORTANT

Don't merge this PR since we are moving to autoware-ml.

What

T4MetricV2's min_num_points GT filter recomputed per-box point counts with
points_in_rbbox on the model input point cloud (multi-sweep concatenation,
after remove_close and range filtering). This PR changes it to use the
annotation num_lidar_pts stored in the info pkl (keyframe count, the
nuScenes/Waymo convention).

Why

Recomputing on the input cloud couples the evaluation GT set to the model input
pipeline:

• Changing the sweep count or any point-cloud preprocessing changes which GT
  boxes are evaluated, so runs with different input configs are not comparable
  (an experiment that improves detection can show a lower mAP simply because its
  input config resurrects harder GT boxes).
• Sweep accumulation is ego-motion compensated but not object-motion
  compensated, so the recomputed counts are physically wrong for moving
  objects (smeared points fall outside the box).
• Preprocessing artifacts (remove_close, range filter) remove points from
  boxes that the sensor actually observed, wrongly excluding them from GT.

The benchmark GT should be a fixed property of the dataset, not of the model
config. Note that the training-side filters (ObjectMinPointsFilter)
intentionally keep recomputing on the input cloud — that is a training
hyperparameter and is not changed here.

Verification

On the BEVFusion-L j6gen2_base val split (3,645 frames, kokseang_2_8_1 infos):

• The num_lidar_pts field is fully populated (209,759 instances, 0 missing;
  17.5 % zero-point boxes from interpolated annotations are correctly dropped
  by min_num_points=2).
• The filtered GT set now matches an independent implementation
  (autoware-ml detection metric suite) class by class
  (car 92,351 / bus 3,755 / bicycle 2,456 / pedestrian 35,186 /
  traffic_cone 3,583 / barrier 524).
• 0–121 m mAP moves 0.6518 → 0.6499 and now agrees with the independent
  implementation to ±0.0004 (was −0.006 before this fix).

Side benefit: removes one points_in_rbbox call per frame from evaluation.