Apply LLM middleware and hooks to final ToolCallLoop call

[loning]

This pull request refactors the LLM invocation logic in ToolCallLoop to improve code reuse and clarity, and updates the loop's behavior when the maximum number of tool call rounds is reached. It also enhances test coverage to ensure the new logic is properly exercised. The most important changes are grouped below.

Refactoring and Code Reuse

• Extracted the LLM invocation logic (including hooks and middleware) from the main loop into a new private method InvokeLlmAsync in ToolCallLoop, reducing duplication and improving maintainability. (src/Aevatar.AI.Core/Tools/ToolCallLoop.cs [1] [2]

Behavior Change: Final LLM Call

• When the maximum number of tool call rounds is exhausted, the loop now makes a final LLM call without tools to ensure a textual response is produced, and appends it to the messages if present. (src/Aevatar.AI.Core/Tools/ToolCallLoop.cs src/Aevatar.AI.Core/Tools/ToolCallLoop.csL135-R157)

Testing and Middleware

• Updated the test ExecuteAsync_WhenMaxRoundsReachedWithoutTerminalContent_Should to verify that the final LLM call is made without tools, and that hooks and middleware are invoked the correct number of times. (test/Aevatar.AI.Tests/ToolCallLoopTests.cs [1] [2]
• Added a DelegateLlmCallMiddleware helper class to facilitate testing of LLM call middleware. (test/Aevatar.AI.Tests/ToolCallLoopTests.cs test/Aevatar.AI.Tests/ToolCallLoopTests.csR284-R289)

[eanzhao]

这次改动的价值比较明确：

• 把 ToolCallLoop 里的 LLM 调用逻辑抽到 InvokeLlmAsync，减少重复代码，hook/middleware 生命周期也更一致；
• maxRounds 用尽后新增一次 不带 tools 的 final LLM call，能提高拿到最终文本回复的概率；
• 测试已覆盖 final call 的关键行为（Tools == null，以及 LLM middleware / hook 被调用次数）。

有两个非阻塞的小建议（低优先级）：

1. 增加一个正向用例：final no-tools call 返回文本时，断言返回值和 messages 中 assistant 文本都正确；
2. 增加一个 terminate 用例：验证 final call 在 middleware Terminate = true 时的短路行为（尤其是 provider 调用次数和返回内容）。

结论：LGTM，可合并（建议后续补上上述两条测试，回归风险会更低）。