SpeedLimit: Neural Architecture Search for Quantized Transformer Models