num_res_blocks as Sequence or int

[Warvito]

Add the option to define num_res_blocks in the AutoencoderKL and DiffusionModelUNet as a sequence of integers, indicating the number of residual blocks per level. This would be inline with what is proposed in https://arxiv.org/abs/2301.11093 (in their case they use Uvit) to scale the network, adding more residual blocks only in the lower resolution levels.