Image Type and support in sem_operators

[harshitgupta412]

Changes:

• Added ImageArray type for storing images
  • loads the image lazily from the stored data. Caution: There are no explicit checks on what is stored; the only way to confirm whether the data is proper is by accessing it.
• Updated sem_ops mentioned in the title to allow multimodal data.
  • Major updates in lotus/templates/task_instructions.py and retriever_models to support more general types.
    • Added df2multimodal_info to load multimodal data properly
    • Clip model can be used through sentence transformer.
    • updated the user prompts to send images properly (user_message_formatter and context_formatter)
    • index and call can now take pd.Series or Images as input.
• Tests in .github/tests/multimodality_tests.py
• Example in examples/op_examples/multimodal_ops