fix(prebuilt): retry create_react_agent on MALFORMED_FUNCTION_CALL finish_reason

[Saakshi Gupta (saakshigupta2002)]

Description: When using create_react_agent with Google Gemini models, the agent silently terminates without making any tool calls when the LLM returns finish_reason: MALFORMED_FUNCTION_CALL. The should_continue conditional edge only checks for tool_calls presence on the AIMessage — when Gemini fails to produce valid JSON for a tool call, it returns an AIMessage with no tool_calls and sets finish_reason: MALFORMED_FUNCTION_CALL in response_metadata. The router sees no tool calls and routes to END, with no indication of the failure.

This PR adds detection of retryable finish_reason values in the should_continue router. When a retryable finish reason is detected (currently MALFORMED_FUNCTION_CALL), the agent logs a warning and routes back to the entrypoint node for a retry attempt. LangGraph's built-in recursion limit (default 25) prevents infinite retry loops.

Key changes:

• Added _RETRYABLE_FINISH_REASONS module-level constant in chat_agent_executor.py — an extensible set of finish reasons that trigger retry
• Modified should_continue to inspect response_metadata.finish_reason when no tool calls are present; retryable reasons route back to the agent entrypoint
• Added entrypoint to agent_paths so the conditional edge accepts the retry target
• Extended FakeToolCallingModel in tests with response_metadata_list support
• Added 4 unit tests covering: retry on malformed call (v1/v2), no-retry on normal finish reason, warning log verification, and retry with pre_model_hook
• Updated graph structure snapshots

Design rationale: retry (instead of raising) aligns with Google's own recommendation for handling MALFORMED_FUNCTION_CALL, and is backwards compatible since the only behavioral change is that previously-broken flows now recover instead of silently failing.

Closes #6574

Issue: #6574

Dependencies: None