pub fn log_gemm_shape(gpu: &dyn GpuBackend, name: &str, m: u32, n: u32, k: u32)
2026-09-25: With METRALE_GEMM_SHAPE_LOG=1, log each (kernel, M, N, K) GEMM shape once per backend at warn level, with its FLOP count.
METRALE_GEMM_SHAPE_LOG=1