Phi-4 support

[filipw]

Phi-4 is now available on HF:

• https://huggingface.co/microsoft/phi-4-gguf
• https://huggingface.co/microsoft/phi-4

When trying to run the Phi-4 GGUF using the existing quantized_phi3 implementation I get:

loaded 243 tensors (9.05GB) in 0.27s
model built
Error: shape mismatch in reshape, lhs: [1, 12, 1280], rhs: [1, 12, 40, 128]
Write a function to count prime numbers up to N. %     

When trying to run the Phi-4 GGUF using the quantized_llama implementation I get:

loaded 243 tensors (9.05GB) in 0.30s
Error: cannot find llama.attention.head_count in metadata

These are reproducible via the quantized-phi examples by just swapping the models with:

"microsoft/phi-4-gguf",
"phi-4-q4.gguf",
"main",

It would be great to have Phi-4 support.

[filipw]

FWIW this is the change in llama.cpp ggml-org/llama.cpp#10817