Text this: Leveraging Simplex Gradient Variance and Bias Reduction for Black-Box Optimization of Noisy and Costly Functions