fix: improved caching by simply keeping groupings of tool calls by terminating a round on each tool merge into the bucket, rather than waiting for LLM narration to end a round.

src/client/vertexai.rs

@@ -600,6 +600,7 @@ mod tests {
                         tool_results: vec![],
                         text: "leftover text".to_string(),
                         sequence: false,
+                        round_starts: vec![],
                     }),
                 ),
             ],