[Christian-Sidak]

Summary

• connection.py: replace deepcopy(flow._context) with dict(flow._context) to give per-run isolation without deep-copying user-supplied values
• flow.py update_context: replace deepcopy(self) with copy.copy(self) and a shallow dict merge, avoiding a full deep-copy of the Flow including callbacks on every call
• model_invoker.py create_with_llm_observe: replace deepcopy(middleware) with list(middleware), passing instances through by reference so stateful middleware accumulates state correctly across LLM calls

The dict(...) shallow copy preserves the existing contract: rt.context.put(...) inside a run writes to a per-run copy of the top-level dict, so it cannot leak into flow._context or the next run.

Test plan

• test_flow_invoke_passes_non_copyable_context_object: threading.Lock in context no longer raises TypeError
• test_flow_invoke_context_value_is_same_object_not_clone: run sees the callers exact object (is check)
• test_flow_update_context_non_copyable_object: update_context with a lock in context no longer raises
• test_flow_update_context_does_not_modify_original: original Flow is unchanged after update_context
• test_flow_context_mutations_isolated_between_runs: rt.context.put inside a run does not bleed into flow._context
• test_create_with_llm_observe_preserves_stateful_middleware: counting middleware accumulates state across two invoke calls (reads 2, not 0)
• test_create_with_llm_observe_non_copyable_middleware_does_not_crash: middleware holding a threading.Lock no longer raises TypeError
• Existing test suite passes: all 30 unit tests and 21 integration flow tests green locally

Fixes #1675