Remove pipeline_state from tensorRT/ — it ships with the inference code repo, not here

Add pipeline_state + per-model padding_embedding for SA3 TRT pipeline

Add SA3 TRT engines (LFS-tracked)

Create tensorRT/README.md