[Model] Define merge_by_field_config MM interface (R-T)

[ayushsatyam146]

Part of #26149

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @ayushsatyam146.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[gemini-code-assist]

Code Review

This pull request introduces a new merge_by_field_config interface for multimodal models to optimize input batching. While the refactoring is a good step, it introduces a critical bug in several models (step3_vl.py, tarsier.py). The changes incorrectly assume that multimodal data like pixel_values will always be a single stacked tensor. However, for batches with heterogeneous input sizes (e.g., images of different dimensions), the data will be a list of tensors. The previous flatten_bn function, which was removed, handled this by concatenating the list. This concatenation logic is now missing, which will lead to AttributeError and crashes. I've added suggestions to restore the concatenation logic in the affected models.

[DarkLight1337]

Have you tested each model using the example script?

[ayushsatyam146]

@DarkLight1337 I have done the required changes. But, I am sorry since I couldn't run the example scripts due to GPU constraints

[DarkLight1337]

Thanks for helping out! However I found that both Step3-VL and Terratorch fail to be run locally. I'll try to fix them later.