fix(streampetr) sampler always skips first element of sequence

[jacoblambert]

Summary

This PR fixes a bug in StreamPETR’s GroupStreamingSampler where the first frame of each sequence was always dropped, despite the sampler intending to preserve it according to the comment:

# Randomly drop some elements from the group based on the random_drop_probability, to simulate camera drop. But never drop the first element because it might cause the value of prev_exists to be wrong in the inputs.

This was a logic fault is because this list comprehension deciding which frame to KEEP, not which frame to drop so

if np.random.rand() > self.random_drop_probability and position > 0

makes it impossible to keep frame0. I changed the condition to match the intention of the comment.

For fusion use cases like in auto_labeling_3d, this caused StreamPETR predictions to become temporally misaligned with lidar frames during ensemble/tracking. The change affects StreamPETR training/inference, but is limited to StreamPETR and does not affect non-StreamPETR pipelines.

[KSeangTan]

LGTM! Thanks for the fix