mtmd: Add DeepSeekOCR Support

[sfallah]

Feature Request: #16676

Make sure to read the contributing guidelines before submitting a PR

[sfallah]

@Acly
I thought it might interest you!
I have implement add_rel_pos_inplace and get_rel_pos (for DeepseekOCR SAM-Vit) without using the cpu only ggml ops (ggml_get_rel_pos etc.) and it works on Metal.
I will also test on CUDA tomorrow, and hopefully the CI checks will later show if the implementation also works for other devices.

You will find the code here:
https://github.com/sfallah/llama.cpp/blob/sf/deepseek-ocr/tools/mtmd/clip.cpp#L2473
https://github.com/sfallah/llama.cpp/blob/sf/deepseek-ocr/tools/mtmd/clip.cpp#L2502

Disclaimer:
My implementation is very rudimentary but @bluebread reviewed and already improved it a bit in separate branch.
I have been using ChatGPT to translate from original pytorch code but have debugged the code and got it to run.
The functions have also been tested with and compared with ggml/examples/sam (sam example from ggml repo), the results align well.

[ngxson]

I have implement add_rel_pos_inplace and get_rel_pos (for DeepseekOCR SAM-Vit) without using the cpu only ggml ops (ggml_get_rel_pos etc.) and it works on Metal.

Just wondering, which ops are currently CPU-only (apart from ggml_arange?)

Beside, it's better introduce new op in a follow-up PR, to avoid adding too much complexity to the review process.