feat: Implement dropout regularization and enhance testing

[kholdrex]

This PR adds complete dropout regularization support to the Rust LSTM library, implementing multiple dropout types specifically designed for recurrent neural networks along with comprehensive configuration options and training integration.

Features Added

🎯 Core Dropout Implementation (src/layers/dropout.rs)

• Standard Dropout: Element-wise random zeroing with configurable rates
• Variational Dropout: Consistent mask across time steps for RNN stability
• Zoneout: LSTM-specific regularization keeping previous hidden/cell states
• Proper forward/backward pass with mask saving for gradient computation
• Training/evaluation mode switching with automatic mask management

🧠 LSTM Integration (src/layers/lstm_cell.rs)

• Input Dropout: Applied to input sequences before gate computation
• Recurrent Dropout: Applied to hidden states in recurrent connections
• Output Dropout: Applied to final hidden state outputs
• Zoneout Integration: Cell and hidden state regularization
• Builder pattern for easy configuration: .with_input_dropout(0.2, true)
• Mathematically correct backward pass with saved dropout masks