[Haralishev77]

Summary

Introduce model-specific V2X communication boundaries so each CAV performs the appropriate local inference stages before transmitting data. Payloads now represent the actual information exchanged by Early, Late, Intermediate, Where2Comm, and FPV-RCNN fusion pipelines.

Related issue

Related to #192

Changes

• Added typed, model-specific V2X payloads and communication adapters.
• Changed Early fusion to transmit sender-local LiDAR and project it at the receiver.
• Changed Late fusion to transmit decoded bounding boxes and confidence scores, followed by receiver-side projection and global NMS.
• Changed Intermediate models to transmit learned spatial or multi-scale features instead of preprocessed voxel inputs.
• Applied model compression before transmission and decompression before receiver-side fusion.
• Applied the Where2Comm communication mask at the sender-side transmission boundary.
• Kept GT, labels, anchors, object IDs, and visualization context local to the dataset.
• Preserved FPV-RCNN cooperation using stage-one proposals and VSA point features without GT-dependent inference.
• Added regression and equivalence tests for communication contracts, metadata, post-processing, compression, packet delivery, and model splits.

Validation

• Simulation run
• Manual verification

Validated Early, Late, Where2Comm, and FPV-RCNN communication paths in simulation. Added unit tests for the new and modified functionality; the complete CI test suite remains to be run.

Checklist

• Linked the related issue
• Kept the PR focused and reviewable