Add integration tests with real models

[MaxGhenis]

Description

Add integration tests using real architectures (ResNet, VGG, Transformer) to ensure L0 works with production models.

Motivation

Unit tests cover individual components, but we need to verify L0 works with complete, real-world architectures.

Test Cases

# tests/integration/test_resnet.py
def test_resnet18_cifar10():
    model = torchvision.models.resnet18(num_classes=10)
    model = add_l0_to_model(model, init_sparsity=0.5)
    
    # Train for 1 epoch on small subset
    train_loader = get_cifar10_subset(100)
    train_one_epoch(model, train_loader)
    
    # Verify:
    # - Forward/backward passes work
    # - Sparsity is being learned
    # - Loss decreases
    
# tests/integration/test_transformer.py  
def test_bert_tiny():
    # Test with tiny BERT model
    ...

Models to Test

• ResNet (18, 50)
• VGG
• MobileNet
• BERT/GPT-2 (tiny versions)
• Simple CNN/RNN