Llama-bench: allow benchmarking lora impact

[IMbackK]

Allows benchmarking a with a lora loaded.

As i dont know how loras are applied exactly in lcpp (esp in the case of quantized weights, where it seams just merging in ram would not be trivial) im not sure if that makes sense to do, so this is a draft.

[slaren]

Yes absolutely, having a way to measure performance with one or more loras applied would be very useful. Loras are not merged into the weights, they are applied during inference, see the function llm_build_lora_mm for an example of how this is done.

[IMbackK]

Great, i will undraft this then.

This pr only allows one lora as the llama-bench permutation interface, ie you can specify --lora none --lora some_file and it will bench first with no lora and then with a lora or --lora a,b as with any bench parameter, makes it a bit hard to come up with a sane way to do 2d permutation with loras.

[slaren]

Yes, I don't think it is very useful to have the loras as part of the test grid. Loras are strongly tied to a model, I don't think they can be separated. I would suggest adding some syntax to -m such as -m model.gguf+lora.gguf+lora2.gguf, then it would be possible to test at the same time e.g. a model, and the same model with a lora applied.

[IMbackK]

Please dont merge as is, i will make improvements as suggested by @slaren