Add visualization tools for sparsity patterns

[MaxGhenis]

Description

Create utilities to visualize which weights/neurons/filters are being pruned by L0.

Motivation

Understanding sparsity patterns helps debug issues and provides insights into what the model considers important.

Proposed Visualizations

from l0.visualization import plot_sparsity

# 1. Layer-wise sparsity
plot_sparsity.by_layer(model)  # Bar chart of sparsity per layer

# 2. Weight heatmaps
plot_sparsity.weight_heatmap(model.fc1)  # Heatmap of gate values

# 3. Sparsity evolution
plot_sparsity.evolution(checkpoint_dir)  # Sparsity over training

# 4. Filter importance (for CNNs)
plot_sparsity.filter_importance(model.conv1)  # Which filters are kept

# 5. Attention head importance (for Transformers)
plot_sparsity.head_importance(model.attention)

Features

• Matplotlib/Plotly backends
• Interactive visualizations
• Export to images/HTML
• Integration with TensorBoard/WandB