perf: speed up format 2.2 300% by spawning structural decode batch tasks

rust/lance-encoding/src/decoder.rs

@@ -1795,7 +1795,15 @@ impl StructuralBatchDecodeStream {
                 let emitted_batch_size_warning = slf.emitted_batch_size_warning.clone();
                 let task = async move {
                     let next_task = next_task?;
-                    async move { next_task.into_batch(emitted_batch_size_warning) }.await
+                    // Real decode work happens inside into_batch, which can block the current
+                    // thread for a long time. By spawning it as a new task, we allow Tokio's
+                    // worker threads to keep making progress.
+                    tokio::spawn(async move { next_task.into_batch(emitted_batch_size_warning) })
+                        .await
+                        .map_err(|err| Error::Wrapped {
+                            error: err.into(),
+                            location: location!(),
+                        })?
                 };
                 (task, num_rows)
             });