pin llama.cpp build version with docker for both platforms

images/benchmark-llm/Dockerfile

@@ -1,15 +1,17 @@
-FROM ghcr.io/ggerganov/llama.cpp:full AS base_cpu
+# when updating the build number, make sure both AMD64 and ARM64 builds are available
+ARG LLAMA_CPP_BUILD=b4524
+FROM ghcr.io/ggerganov/llama.cpp:full-${LLAMA_CPP_BUILD} AS base_cpu
 # collect and copy shared libs for CPU-optimized benchmarks on AMD64,
 # where the default build is CUDA
 COPY extract-shared-cpu-libs.sh /tmp/extract-shared-cpu-libs.sh
 ARG TARGETARCH
 RUN if [ "$TARGETARCH" = "amd64" ]; then /tmp/extract-shared-cpu-libs.sh; fi
 RUN mv /app /llama_cpp_cpu
 
-FROM ghcr.io/ggerganov/llama.cpp:full-cuda AS base_amd64
+FROM ghcr.io/ggerganov/llama.cpp:full-cuda-${LLAMA_CPP_BUILD} AS base_amd64
 RUN mv /app /llama_cpp_gpu
 
-FROM ghcr.io/ggerganov/llama.cpp:full AS base_arm64
+FROM ghcr.io/ggerganov/llama.cpp:full-${LLAMA_CPP_BUILD} AS base_arm64
 
 ARG TARGETARCH
 FROM base_${TARGETARCH} AS final