Convolutional VAE is not flexible to different image sizes.

[renan-siqueira]

Description:

Currently, the ConvolutionalVAE design is tailored to specifically handle images of size 64x64. Due to hardcoded dimensions in the dense (fully connected) layers, the model cannot easily adapt to different image sizes without manual modifications.

Steps to reproduce:

1. Run the project with "conv_vae" type.
2. Try to forward an image of a different size, e.g., 128x128 or 32x32, through the model.
3. Observe the mismatch error or unexpected behavior due to fixed input-output sizes of certain layers, especially the dense layers related to the latent space.

Expected outcome:

The ConvolutionalVAE should ideally be as flexible as the ConvolutionalAutoencoder in handling any square image size.

Current outcome:

The ConvolutionalVAE can only handle images of size 64x64 without raising dimension-related errors or producing unexpected results.