llama-bench Restores Flash Attention Defaults for GPU Acceleration

Build b9437 fixes a critical oversight where llama-bench previously disabled Flash Attention by default, even on GPU-accelerated hardware. This change aligns llama-bench with llama-server and llama-cli by automatically enabling Flash Attention on capable backends (CUDA, Metal, Vulkan), delivering proper GPU acceleration for performance benchmarking. Users with pre-b9437 baselines on GPU hardware should re-run their tests to ensure results accurately reflect actual performance with Flash Attention enabled.