mtmd-cli: reset roofline profiler after warmup prefill

[roberteg16]

Summary

This PR wires the existing per-op HIP roofline profiler into llama-mtmd-cli so VLM prefill can produce a clean, prefill-only artifact. After the multimodal warmup encode pass completes, ggml_cuda_roofline_reset() discards everything recorded during warmup; the subsequent measured prefill (vision encoder + text prompt) is what lands in the GGML_ROOFLINE_OUT report on exit. The change mirrors the post-warmup reset already used in llama-bench for text models.

Motivation

The roofline profiler added in #33 is already usable for text prefill through llama-bench (-p N -n 0 -r 1 with GGML_ROOFLINE_OUT set). VLM models are exercised through llama-mtmd-cli (or llama-server), and mtmd performs its own warmup encode during initialization. Without a reset at the prefill boundary, the atexit report would blend warmup vision-encoder ops with the measured prefill, making the artifact unusable for the prefill roofline KPI in rocm-scripts (tools/generate_llamacpp_ttft_from_artifacts.py expects a prefill-only slice).

What changed

tools/mtmd/mtmd-cli.cpp includes ggml-cuda-roofline.h and calls ggml_cuda_roofline_reset() immediately after the warmup message is evaluated and before the measured prefill eval_message runs. tools/mtmd/CMakeLists.txt defines GGML_HIP_ROOFLINE for the llama-mtmd-cli target and adds the ggml-cuda include path when the HIP roofline option is enabled at build time, matching the llama-bench wiring. Both the include and the reset call are behind #ifdef GGML_HIP_ROOFLINE, so non-roofline builds are unchanged.

Usage

Build with -DGGML_HIP=ON -DGGML_HIP_ROOFLINE=ON, then run: GGML_ROOFLINE_OUT=/tmp/vlm-roofline.json ./build/bin/llama-mtmd-cli -m model.gguf --mmproj mmproj.gguf --image img.jpg -p "describe this" -n 0. The JSON written on exit should cover the measured prefill only (vision encode + text prompt), with warmup ops excluded.