Added reflexion prompting agent from inheriting ReAct agent.

[PalmPalm7]

What does this PR do?

Link to issues: meta-llama/llama-stack#1389

This PR adds the Reflexion Agent, a new agent type that extends ReAct with self-reflection capabilities. The Reflexion Agent can reflect on its past reasoning and errors to improve subsequent responses within a conversation.

Test Plan

I tested the implementation with the following steps:

1. Created a sample Reflexion Agent using the new class:

reflexion_agent = ReflexionAgent(
    client=client,
    model="meta-llama/Llama-3.1-8B-Instruct",
    builtin_toolgroups=["builtin::websearch"],
    client_tools=[calculator_tool]
)

2. Tested multi-turn conversation to verify reflection persistence:

# First query with a deliberate error
response1 = reflexion_agent.create_turn(
    messages=[{"role": "user", "content": "What is 15% of 85?"}],
    session_id=session_id
)
# Response correctly calculated 12.75

# Second query referencing the first
response2 = reflexion_agent.create_turn(
    messages=[{"role": "user", "content": "What if we round to the nearest integer?"}],
    session_id=session_id
)
# Response correctly reflected on previous calculation and rounded to 13

3. Verified that reflections persist between turns by examining the agent's reflection_memory:

assert session_id in reflexion_agent.reflection_memory
assert len(reflexion_agent.reflection_memory[session_id]) > 0

4. Confirmed correct JSON schema validation for the ReflexionOutput format.

5. Tested with both streaming and non-streaming responses to ensure reflection collection works in both cases.