create a script to train autoencoderkl

[lavinal712]

Add AutoencoderKL Training Script

Description

This PR adds a complete training script for AutoencoderKL models. The script supports the following features:

• Multiple loss functions (L1/L2 reconstruction loss, LPIPS perceptual loss, KL divergence)
• Integrated adversarial training
• Mixed precision training support
• Comprehensive training logging and validation
• Hugging Face Hub model upload
• Detailed command-line parameter configuration

Key Features

Core Training Pipeline

• Complete VAE training loop implementation
• Multi-GPU distributed training support
• Gradient accumulation and checkpoint saving

Loss Functions

• Reconstruction loss (L1/L2)
• LPIPS perceptual loss
• KL divergence regularization
• Adversarial loss

Optimization & Performance

• 8-bit Adam optimizer support
• Integrated xFormers memory optimization
• TF32 acceleration support

Monitoring & Validation

• TensorBoard and WandB logging
• Periodic validation with sample image saving
• Detailed training metrics monitoring

@sayakpaul