Fix DenseLayer backpropagation and MNIST download reliability

[Copilot]

Neural network fails to learn on MNIST (~10% accuracy = random guessing) and file downloads intermittently corrupt.

Backpropagation Fix

DenseLayer.backpropagate() was propagating raw error instead of the computed gradient delta:

// Before: bypasses activation derivative
return weights.transpose().multiply(error);

// After: propagates delta correctly  
return weights.transpose().multiply(grad);

The grad already incorporates error × activation_derivative — the signal that must flow through the weight matrix to the previous layer.

Download Fix

downloadFile() used Files.copy() with an unbuffered URL stream, causing truncated GZIP files and EOFException.

• Replaced with explicit buffered I/O (8KB chunks)
• Reads until true EOF
• Maintains replace-if-exists semantics via StandardOpenOption.TRUNCATE_EXISTING

Original prompt

Исправь DenseLayer, нейросеть плохо обучается, после 10 epoch обучения на mnist правильность ответов ~10%

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.