respect -fitc from llama-bench instead of using the benchmark context size

[MartinEmrich]

Overview

While tweaking the optimal settings for running Qwen 3.6 MoE on a small mobile GPU plus System RAM, I was wondering why what llama-fit-params used with llama-bench produced different results that using -fitc with llama-bench directly.

Disclosure: I used AI (Qwen 3.6, of course via llama-server) for the analysis and the fix. These words here are my own though.

Additional information

Turns out that llama-bench does not use my -fitc 196608 parameter (I wanted to benchmark for) for fitting, but strictly the actual benchmark context size (pp512 plus tg128, makes only 640 tokens). All that -fitc does is enable fitting.
I then tried absurdly different -fitc 16384 as well as -fitc 1638400 parameters to verify, and all produced the same benchmark result (higher, as more layers and parameters fit on the GPU).

At the changed code, if n_ctx_auto is false (which is always the case with llama-bench), it now clamps the context size for fit calculation to the passed -fitc parameter via fit_min_ctx if it is bigger (most likely in practice) than the benchmark's token requirements.

This way the benchmark runs with the same settings I get from llama-fit-params with the actual context I plan to run llama-server with.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, for analysis and implementing the change.