A Configurable Floating-Point Fused Multiply-Add Design With Mixed Precision for AI Accelerators