Improve handling of exceptions from tool calls in `Chat`

[jackmpcollins]

The Chat.exec_function_call should optionally be able to catch exceptions raised when executing the FunctionCall and append an appropriate result message containing details of the error. It should still just make a single function call and append a single message, and any required looping/retry should be handled outside of this method.

Worth considering if this can/should be unified with RetryChatModel? RetryChatModel is concerned with handling errors generating or parsing the LLM response, whereas this would handle exceptions raised by the user-provided functions, so maybe makes sense to keep separate.

Related to #417

[jackmpcollins]

• Cover parameter-validation errors (wrong/missing args) and exceptions that surface mid-stream?
• Wrap param errors in a new ToolValidationError, and streamed exceptions in ToolStreamError, then convert both to a ToolResultMessage so RetryChatModel can decide to continue/retry ?
• Add tests: (a) bad params, (b) tool raises, (c) streamed error -> assert chunks are recombined and a single result message is appended.

[CHOS1N11111]

Hi! I'd like to work on this.

My current understanding is:

• Chat.exec_function_call should optionally catch exceptions raised while executing a FunctionCall
• when enabled, it should append a result/error message containing the exception details to the chat
• it should still only execute a single function call and append a single message
• any retry / loop logic should stay outside this method

My plan is to keep the default behavior unchanged and add an opt-in flag for exception catching, with tests covering both the existing raise behavior and the new append-error-message behavior.