mtmd: build_vit batching

[sfallah]

Overview

This PR introduces an optional batch dimension in build_vit, so a
caller can encode several same-size inputs (image tiles, frames) in one graph.
No change for existing models: that means for a 2D [n_embd, n_pos]
input (B == 1), nothing changes.

Changes

• build_vit takes inp as [n_embd, n_pos] or [n_embd, n_pos, B].
• Body runs flattened 2D [n_embd, n_pos * B]; the batch only reappears in
  self-attention as 4D [d_head, n_head, n_pos, B] Q/K/V views. Output restored
  to [n_embd, n_pos, B].

First consumer: DeepSeek-OCR multi-tile encoding (#24300, stacked on this).

Testing

Built llama-mtmd-cli; DeepSeek-OCR single-view still matches (the B == 1 path).
Ran tools/mtmd/tests.sh big;
all tests that pass on master pass here too.
The huge variant is not tested.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - I used AI assistance for code review, debugging, implementation checks, and testing. I have reviewed the submitted changes and take responsibility for the full contents of this PR.

[ngxson]

can you run ./tools/mtmd/tests.sh and report the results here?

note: granite is known to be broken

[sfallah]

can you run ./tools/mtmd/tests.sh and report the results here?

note: granite is known to be broken

I have already, there are two that are failing when I run tools/mtmd/tests.sh big
The exact same two that failed on master (my base).

[vision] FAIL: ibm-research/granite-vision-3.2-2b-GGUF:Q4_K_M
[vision] FAIL: ggml-org/HunyuanVL-4B-GGUF:Q8_0

[sfallah]

@ngxson
FYI: ggml-org/HunyuanVL-4B-GGUF:Q8_0 fails because it doesn't exist on HF hub.

[ngxson]

yes that's expected

[ngxson]

note that batching is not just for multi-tile encode, but it should eventually allow batching multiple images of same size. that will be important for video processing where we need to process multiple images in the same pass

I will fix this comment along with my refactoring to add the proper architecture for doing so