Fix kernel_size from quant_conv and post_quant_conv layers

[Warvito]

There is a typo in the AutoencoderKL post_quant_conv, quant_conv_mu, and quant_conv_log_sigma. These should be conv layers with kernel size 1x1 (or 1x1x1) (as show in here and here). Currently they have kernel size 3x3x3.

[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)