Investigate MNIST data loader functionality

[oulrich1]

Created documentation structure in docs/ directory:

• docs/README.md: Documentation index and standards
• docs/experiments/mnist_experiments.md: Detailed MNIST experiments

Documentation includes:

• Three experiment configurations (full, mini, baseline)
• Network architectures (784→256→128→10, 784→128→10, 784→64→10)
• Activation functions: ReLU (hidden), Sigmoid (output)
• Optimizer: Adam with learning rates 0.001-0.01
• Loss function: Cross-Entropy
• Test results from test suite execution
• Code examples for data loading, network creation, and training
• Reproducibility instructions with build commands
• Expected performance metrics (85-92% accuracy for full training)

Test results documented:

• Network creation: ✓ PASSED
• Batch training: ✓ PASSED (loss 0.5871 → 0.0000)
• Accuracy computation: ✓ PASSED (80% on test data)