Extract features over detections rather than frames in tracker

[aaprasad]

Currently in the tracker we extract the features per frame rather than over all the instances at once. In the GlobalTrackingTransformer, we've refactored it so that we stack the instances and then apply the feature extractor.

Current implementation:

        for frame in frames:
            if frame.has_instances():
                if not self.use_vis_feats:
                    for instance in frame.instances:
                        instance.features = torch.zeros(1, model.d_model)
                    # frame["features"] = torch.randn(
                    #     num_frame_instances, self.model.d_model
                    # )

                # comment out to turn encoder off

                # Assuming the encoder is already trained or train encoder jointly.
                elif not frame.has_features():
                    with torch.no_grad():
                        crops = frame.get_crops()
                        z = model.visual_encoder(crops)

                        for i, z_i in enumerate(z):
                            frame.instances[i].features = z_i

Proposed implementation:

if not self.use_vis_feats:
     for instance in frame.instances:
           instance.features = torch.zeros(1, model.d_model)
else:
     instances = [instance for instance in frame.instances for frame in frames]
     features = self.model.extract_features(instances)