Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers