feat: Add learning rate scheduling system

[kholdrex]

This PR adds a comprehensive learning rate scheduling system to the rust-lstm library, bringing it to feature parity with PyTorch and TensorFlow for adaptive learning rate management.

Why This Matters

Learning rate scheduling is critical for training quality in modern deep learning:

• Improves convergence speed and final model performance
• Prevents overfitting and training instability
• Enables state-of-the-art training techniques like OneCycleLR
• Essential for production-ready deep learning workflows

What's New

Core Features

• 8 Different Schedulers: StepLR, MultiStepLR, ExponentialLR, CosineAnnealingLR, CosineAnnealingWarmRestarts, OneCycleLR, ReduceLROnPlateau, LinearLR
• ScheduledOptimizer: Seamless integration between any optimizer and scheduler
• ScheduledLSTMTrainer: Automatic scheduler stepping during training
• Enhanced Metrics: Learning rate tracking in training history

Convenience Functions

// Quick setup with popular combinations
let trainer = create_one_cycle_trainer(network, 0.1, 100);
let trainer = create_step_lr_trainer(network, 0.01, 10, 0.5);
let trainer = create_cosine_annealing_trainer(network, 0.01, 50, 1e-6);

Advanced Manual Configuration

let scheduled_optimizer = ScheduledOptimizer::new(
    Adam::new(0.001),
    OneCycleLR::new(0.1, 100),
    0.001
);