Add multi-points input, foreground/background points input and box input to EfficientSAM model

[Zhang-Yang-Sustech]

The EfficientSAM model in the model zoo can only support single point input to infer before. This time I change the model input dimension to 6, and add a dynamic supplement method to support multi-points input. Users can input 6 points to select segment object at most. Additionally, I select from the output 3 masks to avoid to pick the mask with background points. I also add foreground/background points and box input in the demo.

1. new model(image_segmentation_efficientsam_ti_2024may.onnx) with new input dimension:
   'batched_images': (1, 3, 1024, 1024), previous is (1, 3, 640, 640);
   'batched_point_coords':(1, 1, 6, 2), previous is (1, 1, 1, 2), can only support single point input, now can support 6 points;
   'batched_point_labels':(1, 1, 6), previous is (1, 1, 1), only one label, now each point one label.
   with new output dimension:
   'output_masks': (1, 1, 3, 1024, 1024), previous is (1, 1, 1, 640, 640), can only output single mask, now output three mask, we can select the mask we want to enhance performance;
   'iou_predictions': (1, 1, 3), new output which is used to select the masks by scores.
2. new model file(efficientSAM.py), with dynamical supplement points method to support multi-points input in preprocess, with masks selection method in postprocess.
3. new demo file(demo.py), with foreground/background points input and box input.