This pull request introduces support for multiple learning rate schedulers in the OptimizerLightningModule and OptimizerConfig classes, allowing users to configure and combine different learning rate schedules (e.g., step-wise warmup and epoch-wise plateau) in PyTorch Lightning workflows. The update adds a new LRSchedulerConfig class, deprecates the old single-scheduler interface, and updates documentation and examples to reflect these changes.

Core API Enhancements:

• Added the LRSchedulerConfig dataclass to encapsulate configuration for individual learning rate schedulers, including generator, interval, monitor, frequency, and other options. (aiaccel/torch/lightning/opt_lightning_module.py)
• Updated OptimizerConfig to accept a list of LRSchedulerConfig instances via the new schedulers argument, supporting multiple schedulers per optimizer. The old scheduler_generator/scheduler_interval/scheduler_monitor fields are now deprecated in favor of this new approach, and validation ensures only one method is used. (aiaccel/torch/lightning/opt_lightning_module.py) [1] [2]
• Modified OptimizerLightningModule.configure_optimizers() to return optimizer and scheduler(s) in the format expected by Lightning, supporting multiple schedulers if provided. (aiaccel/torch/lightning/opt_lightning_module.py)
• Updated the __init__.py exports to include LRSchedulerConfig in the public API. (aiaccel/torch/lightning/__init__.py)

Documentation and Example Updates:

• Expanded the user guide and API reference to document the new multiple-scheduler feature, including a YAML configuration example demonstrating how to combine different scheduler types. (docs/source/user_guide/torch.rst, docs/source/api_reference/torch.rst) [1] [2]
• Updated the ResNet50 CIFAR10 example configuration to use the new schedulers field with two schedulers: LinearLR (step) and CosineAnnealingLR (epoch). (examples/torch/image_classification/recipes/resnet50.cifar10/config.yaml)