fix(workflow): deep copy initial_state to prevent mutation leaks across runs

[gauravSsinha]

Description

Fixes AgentWorkflow leaking initial_state mutations across run() calls when the same workflow instance is reused.

Problem

• Multi-agent path (AgentWorkflow): stores self.initial_state in the context by direct reference, so in-place mutations modify the workflow's own state
• Single-agent path (BaseWorkflowAgent): uses .copy() (shallow copy), which avoids top-level leakage but still leaks nested mutable values (lists, dicts)

This violates the documented behavior that workflows are stateless between runs unless a Context is explicitly reused.

Fix

Use copy.deepcopy() in both paths to ensure each run() gets a fully independent copy of initial_state.

Changes

• multi_agent_workflow.py: copy.deepcopy(self.initial_state) instead of self.initial_state
• base_agent.py: copy.deepcopy(self.initial_state) instead of self.initial_state.copy()

Reproduction (from issue)

workflow = AgentWorkflow.from_tools_or_functions(
    [bump_counter], llm=MockFunctionCallingLLM(),
    initial_state={"counter": 0},
)
await workflow.run(user_msg="bump it")  # counter -> 1
state = await (await workflow.run(user_msg="bump it")).ctx.store.get("state")
# Before fix: {'counter': 2}  (leaked from first run)
# After fix:  {'counter': 1}  (independent copy)

Fixes #21774