allow fewshots for multimodal tasks

[artemorloff]

till now only zero-shot tasks may bne processed

if any of the few-shots include images, they are not passed into the model. This behavior causes errors

I suggest processing the images while constructing the task. Fewshot sampler generates labeled_examples and multimodal_args. For future multimodal_args may include any modality. Now only images

[artemorloff]

@baberabb @haileyschoelkopf @lintangsutawika tests fail due to changes in fewshot_context function
it now returns not only ctx, but also multimodal_args - a dict with all files the come with these fewshots

I have checked the code on mmmu (qwen2-vl-2b) and mmlu (to ensure text-only tasks are not affected). Do not see any difference in the output

[artemorloff]

@baberabb @haileyschoelkopf if you like the general idea, i could modify the tests