[Fix] Several Fixes for Multi-View 3D Grounding Baseline

[mxh1999]

Modification

Several fixes for multi-view 3D visual grounding baseline:

• Augmentation: Fix the rotation and scaling behavior of EulerInstance3DBoxes (related to orientations), also affect MV-3DDet baseline
• Data: Fix some data samples with inaccurate tokens_positive corresponding to the target in the language prompt
• Model: Fix create_positive_maps for training with batch_size > 1
• Enhancement: Support training with multiple grounding boxes in a single language prompt

All the fixes will result in a little performance improvement:

Method	[email protected]	[email protected]
Previous	35.15	12.75
Current	36.11	13.43