Refactor error handling and fix state persistence in LlmAgent

[dewitt]

This PR refactors error handling and state persistence in LlmAgent to support robust Human-in-the-Loop (HITL) and client-side tool execution flows.

Key Changes:

• Robust Error Handling: Updates runAndHandleError to support native Error objects and ensures all exceptions (including those in post-processing or function execution) are caught and yielded as standardized error events.
• State Persistence: Fixes a bug in generateRequestConfirmationEvent to ensure actions (including stateDelta) are correctly persisted when the agent suspends for tool confirmation.
• Explicit Suspension: Implements logic to respect pauseOnToolCalls in the agent loop, setting endInvocation = true to return control to the host immediately after a tool call or confirmation request.
• Refactored Testing: Changes callLlmAsync and runAndHandleError to protected visibility and uses a TestLlmAgent subclass to enable strictly typed unit testing without any casts.

Reference: docs/adk-ts-improvements.md

Fixes #143.