feat: Add GRU (Gated Recurrent Unit) implementation with multi-layer support and dropout

[kholdrex]

This PR adds a complete GRU (Gated Recurrent Unit) implementation to the Rust LSTM library, providing an alternative RNN architecture that's simpler yet often as effective as LSTM for many sequence modeling tasks.

✨ Features Added

Core Components

• GRUCell: Complete GRU cell with mathematically correct forward/backward passes
• GRUNetwork: Multi-layer GRU networks with flexible architecture
• Dropout Support: All dropout types (input, recurrent, output, variational, layer-specific)
• Training Integration: Full compatibility with existing optimizers (SGD, Adam, RMSprop)

Key Implementation Details

• ✅ 3-gate architecture: Reset gate, Update gate, New gate computations
• ✅ Efficient caching: Forward pass caching for training efficiency
• ✅ Gradient computation: Accurate backpropagation through time (BPTT)
• ✅ Mode switching: Proper training/evaluation mode handling
• ✅ Memory efficient: Optimized array operations with ndarray

🚀 Benefits

1. Simpler Architecture: 2 gates vs LSTM's 3 gates = fewer parameters & faster training
2. Comparable Performance: Often matches LSTM performance with less complexity
3. Better Gradient Flow: Effective handling of vanishing gradient problem
4. Computational Efficiency: Lower memory and computational overhead
5. Versatility: Excellent for many sequence modeling tasks