refactor: Keep explicit module imports + optimize CNN milestone

Import Strategy:
- Keep explicit 'from tinytorch.core.spatial import Conv2d'
- Maps directly to module structure (Module 09 → core.spatial)
- Better for education: students see exactly where each concept lives
- Removed redundant tinytorch/nn.py (nn/ directory already exists)

Milestone 04 Optimizations:
- Reduced epochs: 50 → 20 (explicit loops are slow!)
- Print progress every 5 epochs (instead of 10)
- Load from local npz file (no sklearn dependency)
- Still achieves ~80%+ accuracy

Educational Rationale:
TinyTorch uses explicit imports to show module structure:
  tinytorch.core.tensor      # Module 01
  tinytorch.core.layers      # Module 03
  tinytorch.core.spatial     # Module 09
  tinytorch.core.losses      # Module 04

PyTorch's torch.nn is convenient but pedagogically unclear.
Our approach: clarity over convenience!

milestones/04_cnn_revolution_1998/cnn_digits.py

@@ -43,7 +43,6 @@ sys.path.insert(0, os.path.join(os.path.dirname(__file__), '../..'))
 from tinytorch import Tensor, SGD, CrossEntropyLoss, enable_autograd
 from tinytorch.core.spatial import Conv2d, MaxPool2d
 from tinytorch.core.layers import Linear, ReLU
-from tinytorch.core.activations import Sigmoid
 from tinytorch.data.loader import DataLoader, TensorDataset
 
 console = Console()
@@ -58,18 +57,17 @@ enable_autograd()
 
 def load_digits_dataset():
     """
-    Load the sklearn 8x8 digits dataset.
+    Load the 8x8 digits dataset from local file.
     
     Returns 1,797 grayscale images of handwritten digits (0-9).
     Each image is 8×8 pixels, perfect for quick CNN demonstrations.
     """
-    from sklearn.datasets import load_digits
+    # Load from the local data file (same as MLP milestone uses)
+    data_path = os.path.join(os.path.dirname(__file__), '../03_mlp_revival_1986/data/digits_8x8.npz')
+    data = np.load(data_path)
     
-    digits = load_digits()
-    
-    # Normalize to [0, 1]
-    images = digits.images / 16.0  # Original range is [0, 16]
-    labels = digits.target
+    images = data['images']  # (1797, 8, 8)
+    labels = data['labels']  # (1797,)
     
     # Split into train/test (80/20)
     n_train = int(0.8 * len(images))
@@ -270,7 +268,7 @@ def train_cnn():
     
     # Hyperparameters
     console.print("\n[bold]⚙️  Training Configuration:[/bold]")
-    epochs = 50
+    epochs = 20  # Reduced for demo speed (explicit loops are slow!)
     batch_size = 32
     learning_rate = 0.01
     
@@ -313,7 +311,7 @@ def train_cnn():
         history["loss"].append(avg_loss)
         history["accuracy"].append(accuracy)
         
-        if (epoch + 1) % 10 == 0:
+        if (epoch + 1) % 5 == 0:  # Print every 5 epochs
             console.print(f"Epoch {epoch+1:3d}/{epochs}  Loss: {avg_loss:.4f}  Accuracy: {accuracy:.1f}%")
     
     training_time = time.time() - start_time

tinytorch/nn/__init__.py

@@ -34,8 +34,9 @@ while this infrastructure provides the clean API they expect from PyTorch.
 """
 
 # Import layers from core (these contain the student implementations)  
-from ..core.layers import Linear, Module  # Use the same Module class as layers
-from ..core.spatial import Conv2d
+from ..core.layers import Linear, ReLU, Dropout
+from ..core.activations import Sigmoid
+from ..core.spatial import Conv2d, MaxPool2d, AvgPool2d
 
 # Import transformer components
 from ..core.embeddings import Embedding, PositionalEncoding