llama: end-to-end tests (#19802)

* tests: add end-to-end tests per model architecture

* fixup for rebase

* fix use-after-free in llama-model-loader.cpp

* fix CI

* fix WebGPU

* fix CI

* disable CI for macOS-latest-cmake-arm64

* use expert_weights_scale only if != 0.0f

* comments

scripts/git-bisect-run.sh

@@ -0,0 +1,18 @@
+#!/usr/bin/env bash
+
+cmake_args=()
+llama_results_args=()
+
+for arg in "${@}"; do
+    if [[ "$arg" == -D* ]]; then
+        cmake_args+=("$arg")
+    else
+        llama_results_args+=("$arg")
+    fi
+done
+
+dir="build-bisect"
+rm -rf ${dir} > /dev/null
+cmake -B ${dir} -S . ${cmake_args} > /dev/null
+cmake --build ${dir} -t llama-results -j $(nproc) > /dev/null
+${dir}/bin/llama-results "${llama_results_args[@]}"

scripts/git-bisect.sh

@@ -0,0 +1,19 @@
+#!/usr/bin/env bash
+
+if [ $# -lt 2 ]; then
+    echo "usage: ./scripts/git-bisect.sh <commit_bad> <commit_good> [additional arguments]"
+    echo "  additional arguments: passed to CMake if they start with \"-D\", to llama-results otherwise"
+    exit 1
+fi
+
+set -e
+set -x
+
+commit_bad=$1
+commit_good=$2
+script_dir="$( cd "$( dirname "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
+git checkout ${commit_good}
+${script_dir}/git-bisect-run.sh --output results.gguf "${@:3}"
+git bisect start ${commit_bad} ${commit_good}
+git bisect run ${script_dir}/git-bisect-run.sh --output results.gguf --check "${@:3}"
+git bisect reset