Compatibility Update for PyTorch-2.3.0

• Updated the code for compatibility with PyTorch 2.3.0 and PyTorch Lightning 2.2.3.
• torchmetrics should still be 0.11.4.
• pydantic-2.7.1 has compatibility issues with lightly-1.4+ and needs to be downgraded if pydantic-2.7.1 is already installed. The error will pop out when installing. Downgrading pydantic does not seem to cause any compatibility issues with other libraries.
• pyYAML-6.0.1 was installed too as an older version was causing an error when installing Pytorch Lightning.
• Trainer.add_argparse_args() has been deprecated in PyTorch-2.0+. However, adding all the args manually seemed too cumbersome. Hence, added add_argparse_args and other relevant functions from an old PyTorch source code in ldm/plargparse.py and ldm/plparsing.py, and used add_argparse_args(Trainer, parser) instead. Works the same.
• Finally, users should change the strategy to ddp or whatever suits best for them. The current code has been set up to run on a single GPU. Line 525 in main.py has been commented out for the purpose.
• PyTorch Lightning-2.2.3 does not support optimizer_idx in training_step. Hence, serialized the execution of optimizer operations in training_step as mentioned in the docs. Also, self.automatic_optimization needs to be set to False.
• accumulate_grad_batches in pl.Trainer is not supported when self.automatic_optimization is set to False. Hence, added a provision for gradient accumulation manually in the training_step.
• Replaced deprecated TestTubeLogger with TensorBoardLogger but forgot to change the name. It still creates a folder named testtube inside the log directory.