fix: use actual tensor embedding dimension instead of model parameter

[chrismuzyn]

Replace hardcoded n_embd with n_embd_tensor in embedding extraction to handle cases where tensor dimensions differ from model parameters.

This issue was only triggered when the pool type was changed. This fixes #18737.

I have tested this fix extensively with Qwen3-VL-Embedding-8b with pooling set to last. Of course, previously it didn't work at all so there was nothing to compare to.

So I also tested extensively with llama-embed-nemotron-8b with pooling set to mean both with and without my change. The returned embeddings matched 100%. (They matched when genned on the same backend, CPU - CPU and Vulkan - Vulkan, they interestingly did NOT match CPU - Vulkan, but I think that must have been the case previously? Unfortunately I somehow broke my ROCM install so I was unable to retest HIP).

Please let me know if there are additional tests I can run or if I can provide my exact GGUFs. Will try to retest HIP if I can get it working again, but the issue was reproducible solely with the CPU backend.

[ngxson]

we recently added n_embd_out for this exact purpose, it should be used instead.

consider adding a GGUF metadata for it. the original Qwen3-VL-Embedding model is also missing 1_Pooling, I don't think it's actually ready to be used unless Qwen team fixed it (I already reached out to them, but got no responses)

[vincenthawke]

we recently added n_embd_out for this exact purpose, it should be used instead.

I am guessing the windows builds don't include that update. Don't see it after running llama-server --help