feat: add transformer head pruning to Pruner

[nathanhubens]

Summary

• Add head_pruning_ratio, prune_num_heads, prune_head_dims parameters to Pruner for structured attention head removal
• Refactor attention detection: _detect_attention_heads() populates num_heads dict without ignoring QKV layers when head pruning is active
• Add _sync_attention_attrs() that reads from torch-pruning's live num_heads state (fixes stale-snapshot bug in old restore_attention_layers)
• Auto-enable XOR pattern: head_pruning_ratio > 0 → prune_num_heads=True, prune_head_dims=False (matching torch-pruning convention)
• Fix PruneCallback.before_fit to forward example_inputs to Pruner (was falling back to default 224×224, breaking transformer model tracing)
• Update print_sparsity() to report attention head count changes

Supported architectures

• timm-style attention (modules with .qkv Linear + .num_heads): full head pruning support
• nn.MultiheadAttention: continues to be ignored (torch-pruning's head pruning requires .out_features on QKV layers, which native MHA lacks — uses raw in_proj_weight parameter)

Usage

# Standalone
pruner = Pruner(model, pruning_ratio=0.3, context='local', criteria=large_final,
                example_inputs=x, head_pruning_ratio=0.5)
pruner.prune_model()

# With PruneCallback
cb = PruneCallback(pruning_ratio=30, schedule=agp, context='global',
                   criteria=large_final, head_pruning_ratio=0.5)
learn.fit(10, cbs=[cb])

Test plan

• Unit tests: head detection, auto-enable XOR, ratio normalization, backward compat (CNN unchanged)
• Slow integration test: head pruning on 4-layer ViT (8→4 heads, param reduction verified)
• Full nbdev-test suite passes
• Pre-push validator: nbdev-export + nbdev-clean + clean checkout verified