Implement mini-batch gradient descent for faster training

[winpat]

This commit introduces mini-batch training to significantly improve
training performance. Instead of updating weights after each sample,
gradients are accumulated over a batch and applied once per batch.

Changes:

• Added gradient accumulation to Linear layer

  • New zeroGradients() method to reset gradients at batch start
  • Modified backward() to accumulate gradients instead of overwriting
  • Updated applyGradients() to average gradients by batch size

• Updated Network training loop

  • Added batch_size parameter to train()
  • Process data in configurable mini-batches
  • Zero gradients at start of each batch
  • Accumulate gradients over batch samples
  • Apply averaged gradients once per batch

• Updated all train() calls to include batch_size parameter

  • main.zig: Uses batch_size=32 for Iris dataset
  • net.zig test: Uses batch_size=1 for backward compatibility
  • README.md: Updated example with batch_size=32

Benefits:

• Faster training through reduced weight update overhead
• More stable gradient estimates from batch averaging
• Better convergence properties
• Configurable batch size for different datasets

The default batch_size of 32 provides a good balance between
speed and gradient stability for most datasets.