llama-fit-params: keep explicit --ctx-size 0

[JohannesGaessler]

Fixes #18376 , alternative to #18567 . If the user explicitly sets --ctx-size 0, set the minimum context size to UINT32_MAX to prevent llama_params_fit from reducing it.

[ggerganov]

@JohannesGaessler On a related note: for LlamaBarn we need a mechanism to query how much memory would be needed for a context of size -c N. Would you accept if we extend llama-fit-params to provide such information and if yes, can you recommend a way to implement it in terms of llama-fit-params UX and libllama API modifications (if needed)?

[JohannesGaessler]

As of right now llama_params_fit internally sets llama_model_params::no_alloc = True, creates llama_model and llama_context instances, and then calls llama_context::memory_breakdown. I would suggest that rather than touch llama_params_fit an external project should simply retrieve this information itself. The problem is that currently the return format of llama_context::memory_breakdown is std::map<ggml_backend_buffer_type_t, llama_memory_breakdown_data> which is not C-compatible. So that functionality needs to either be exposed in a C-compatible way or it needs to be exposed in llama-cpp.h