Storing FFN activation type in GGUF to reduce per-architecture boilerplate

[nisparks]

Ever since first contributing to this project, when Phi-4 came out, I've been a little baffled at why there is so much boilerplate when all the information is present in the HF config files and such. I realized at the time that it is partly due to the conversion from safetensors to gguf, that we lose a bit of information, which in turn gets hardcoded in llama.cpp models.

As someone that has been using AI for some time to develop quickly (I'm guilty), I was putting my tokens to use on this repo, trying to automate this process so we can quickly add support for new architectures. One area that was found lacking was with hidden activation functions. The proposal was to add this metadata to the GGUF file so we don't need as much hand coded architecture when building the model graph.

Per the guides, I'm starting this conversation now, to see if there is support for this, before I submit a relatively simple PR adding this feature. Obviously, I will ensure we are still backwards compatible with existing GGUF files.

Example where this would help: #19888

[nisparks]

Had the AI Agent run an additional audit on GGUF:

FFN Activation (silu, gelu, relu, relu_sqr, etc)
FFN layout (gated vs sequential)
QK-norm type
Post-norm positions (pre, post, pre+post)
Embedding scaling mode (sqrt_embd, none)

However, these additional ones are necessarily needed today.
If ffn_gate tensor exists → gated (PAR). If not → sequential (SEQ).
If attn_q_norm tensor exists → QK-norm enabled.
If attn_post_norm / ffn_post_norm tensors exist → post-norm enabled.
Already covered by f_embedding_scale float in hparams. If > 0, scale is applied.