feat: improve llava next pooling for granite vision

[drbh]

This PR improve LlavaNextForConditionalGeneration to handle models that specify pooled vision_feature_layer.

more concretely the ibm-granite/granite-vision-3.2-2b model specifies

    "vision_feature_layer": [
        -24,
        -20,
        -12,
        -1
    ],

and requires that those layers are pooled. This change allows vision_feature_layer to be a list an iterates over the layers.