Changes

This PR fixes the differences in vision feature we were seeing. Because of the layout, pixtral was getting considered as causal model and thus is_causal was getting set to True for doing scaled_dot_product_attention, whereas we are only using it as encoder for getting the image features in mistral3.py. Therefore, to resolve the issue, we are explicitly setting the attn_name=sdpa_bidirectional when calling vision_tower in get_image_feature function.

This makes the output of FMS look pretty similar to that of using transformers directly.