No convenient way to define tools executed on client side

### Description

When using Fantasy on the server side, there is currently no convenient way to define a tool that should be executed on the client side.
The issue is that when creating a tool via NewAgentTool, the return value of the execution function is always treated as a tool result. However, for client-side tools, we typically don't want to return a tool result from the server at all — we simply want to interrupt the current turn and hand control back to the client so it can execute the tool locally.

It would be best to introduce a clear convention for the execution function’s return value, so this situation can be handled directly inside the tool execution function itself. For example, returning a specific sentinel value (such as YieldToClient) could signal that the tool should be executed on the client and the current turn should be interrupted.