Fixed

• Fix self attention optimization not being triggered, broken residual path [#119]
• Improve speed by not using contiguous Tensors when not needed [#119]

Added

• Attention mask wrapper [#113]
• ViT comparison benchmark [#117]