Crop and pad in LatentDiffusionInferer

[virginiafdez]

To maximise the shape that can go through the VAE or VQ-VAE, sometimes we must pick size shapes that result in latent space shapes that cannot go through the LDM for not being multiple of 2**(num levels on the unet).
To overcome this, a solution is to pad the VAE latent space and crop it back after sampling before passing it to the VAE - VQVAE. Having a simple MONAI transform on the inferer would be enough.