fix: add tool request error feedback to llm

[baduyne]

This PR addresses three related tool request parsing issues:

• TypeError: tool.before_execution() argument after ** must be a mapping, not str #1225 : Prevents tool.before_execution(**tool_args) from crashing when tool_args is returned as a string instead of a dictionary.
• [PROPOSAL] Defensive Parsing for tool_args to Support More LLM Providers #1236 : Adds defensive parsing for tool_args when LLM providers return it as a stringified JSON object.
• Bug: ValueError crash when LLM outputs "args" instead of "tool_args" #1512 : Supports non-canonical tool argument keys by normalizing args to tool_args before validation.
• error：A0: Message misformat, no valid tool request found. #1234: Different LLM providers and models may return tool-calling arguments in slightly different formats, such as args, arguments, or stringified JSON in tool_args.

Solution:
If the runtime cannot safely parse or normalize the tool arguments, it now creates a structured error message and feeds it back into the LLM context. This gives the model clear feedback about what was wrong and guides it to retry the tool call with correctly formatted arguments instead of crashing the tool execution flow.