Make possible to specify attention_head_dim as Sequence

[Warvito]

Since we can have different number of channels across the DiffusionModelUNet, it is important to allow specify in a similar way the number of attention_head_dim.

[Warvito]

Necessary to convert the pretrained model from "Brain Imaging Generation with Latent Diffusion Models" (https://arxiv.org/abs/2209.07162) to this package Unet model (related to #30)