Fix CelebA preprocessing normalization mismatch

[nagateja-naidu]

#Overview

This PR mainly delas with the normalization mismatch as reported in the issue #108 .

###Problem

The VanillaVAE decoder uses nn.Tanh(), which produces reconstructed images in the range [-1, 1]. However, the CelebA preprocessing pipeline only applies transforms.ToTensor(), resulting in input images in the range [0, 1]. Since the reconstruction loss is computed between the reconstructed output and the input tensor using, the input and output ranges are different.

Solution

This PR adds:

transforms.Normalize(
    (0.5, 0.5, 0.5),
    (0.5, 0.5, 0.5)
) to both training and validation processing pipeline such that the input tensor and reconstruction output ranges are same.