perf: replace deepcopy with shallow copy in graph/swarm state management

[AnnasMazhar]

Closes #2275

Summary

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

Motivation

deepcopy on a typical conversation (20 messages, 10 tools) costs 0.218ms per call. In a 5-node graph that's 1.1ms of pure copy overhead. See #2275 for full benchmark.

Changes

• Added _copy_messages() helper: [msg.copy() for msg in messages] (~67x faster)
• Added _copy_model_state() helper: shallow dict copy + tools list copy (~70x faster)
• Replaced 8 copy.deepcopy() calls across graph.py and swarm.py
• Added 2 tests verifying copy isolation (append safety + mutation safety)

Safety

Safe because the executor only:

1. Appends new messages (agent.py:1123) — never mutates existing message dicts
2. Replaces model state values — never mutates nested tool specs

Testing

• hatch fmt --formatter ✓
• hatch fmt --linter ✓ (ruff + mypy pass)
• New tests: test_copy_messages_isolation, test_copy_model_state_isolation

[AnnasMazhar]

Closing this — after deeper analysis I found that the executor DOES mutate existing message dicts in-place:

• agent.py:933: self.messages[-1]["content"] = ... (replaces content list)
• bedrock.py:398: messages[last_user_idx]["content"].append(...) (appends to content)
• repository_session_manager.py:303: messages[index + 1]["content"].extend(...)

A shallow copy (msg.copy()) shares the content list reference, so these mutations would leak back to the saved initial state. deepcopy is necessary here for correctness.

A safe optimization would need to deep-copy only the content lists (not the entire recursive structure), but that requires more careful analysis of all mutation paths. Leaving the issue open for discussion.

Sorry for the premature PR.