Prevent duplicate Deep Agents messages after continue-as-new

[1fanwang]

Long-running Deep Agents workflows can use run_deep_agent to carry conversation
state through continue-as-new. Before this fix, each new run appended the original
input messages to the saved history, so the initial user prompt appeared again after
every rollover. The workflow now resumes from the saved message list once.

State restoration replaces the input's messages field with the snapshot while
preserving the other input fields.

Testing

The regression test starts a real Temporal dev server and forces two
continue-as-new boundaries.

Before

$ git checkout upstream/main -- temporalio/contrib/deepagents/workflow.py
$ uv run pytest tests/contrib/deepagents/test_continue_as_new.py::test_can_threshold_and_cache -vv --tb=short --no-header
E   AssertionError: {'messages': ['start', 'step', 'start', 'step', 'start', 'step', ...], 'todos': [{'content': 'work', 'status': 'completed'}]}
E   assert ['start', 'step', 'start', 'step', 'start', 'step', 'start', 'step'] == ['start', 'step', 'step', 'step']
Results (3.99s):
         1 failed

After

$ git checkout HEAD -- temporalio/contrib/deepagents/workflow.py
$ uv run pytest tests/contrib/deepagents/test_continue_as_new.py::test_can_threshold_and_cache -q --no-header
.                                                                        [100%]
Results (4.15s):
         1 passed