Handled mixed modality

[mano3-1]

This PR introduces support for datasets containing mixed modalities by implementing the following changes:

• Conditional Image Data Inclusion: Checks if there is atleast one image associated with a given datapoint. If no images are tagged, all_pixel_values and all_image_grid_thw are excluded from the final data dictionary, data_dict.

• Homogeneous Batch Sampling: Implements a custom sampler, HomogeneousBatchSampler, to ensure each batch is homogeneous—either containing only text or exclusively text-image pairs.

• Sampler Methods for Training and Evaluation: Adds _get_eval_sampler and _get_train_sampler methods to QwenTrainer

[2U1]

This makes an error of

[rank2]:     sources = self.list_data_dict[i]
[rank2]: TypeError: list indices must be integers or slices, not list

Also, to the best of my knowledge the sampler should just boost the training speed, not a ciritcal issue for running the code. I have some other issues with my env so, I can't really dig in to the code.
I'll try to benchmark the grouping modality code from LLaVA.