perf: replace deepcopy with shallow copy in graph/swarm (~600x faster state management)

[AnnasMazhar]

Summary

Replace copy.deepcopy() with targeted shallow copies in GraphNode and SwarmNode state save/restore operations.

Why

deepcopy recursively copies every nested object. For a typical conversation (20 messages, 10 tools), this costs 0.218ms per call. In a 5-node graph, that's 1.1ms of pure copy overhead per execution.

The executor only:

• Appends new messages to the list
• Replaces content blocks (doesn't mutate existing ones in-place)

This means a shallow copy of the list + shallow copy of each message dict is semantically equivalent but ~600x faster.

Benchmark

Conversation: 20 messages, 10 tools
─────────────────────────────────────────────────
deepcopy (before):    0.218ms per call
shallow copy (after): 0.0003ms per call
Speedup:              ~600x
─────────────────────────────────────────────────

5-node graph execution:
  Before: 1.1ms copy overhead
  After:  0.002ms copy overhead

20-node workflow:
  Before: 9ms latency from copies alone
  After:  0.01ms

Changes

• Added _copy_messages() and _copy_model_state() helpers in graph.py
• Replaced all copy.deepcopy() calls in graph.py and swarm.py
• Imported helpers in swarm.py

Safety

The shallow copy is safe because:

1. agent.py:1123 only calls self.messages.append(message) — never mutates existing messages
2. Model state values are scalars or replaced wholesale
3. The tools list is referenced, not mutated in-place

If a future change mutates message dicts in-place, this would need to be reverted. Added docstrings explaining the invariant.

[AnnasMazhar]

Closing this to follow the contribution guidelines properly. Will open an issue first, then resubmit with hatch fmt/lint verification.