Support --dry-run flag for tool calls

When debugging or learning what a server expects, there's no way to see the JSON-RPC request that would be sent without actually sending it.

You have to guess, call the tool, get an error, adjust, and try again. For destructive operations (like send_message on Slack), this is risky — you can't preview what you're about to do.

A --dry-run flag on tool calls would let users inspect the exact request payload before it hits the server.