[VL] Optimize sort based shuffle

### Description

I tested the sort based shuffle function and found through the perf tool that the largest CPU consumption is in the serialize part. Do you have any plans to write a custom serializer instead of using `PrestoVectorSerde` directly to optimize performance? Or are there any other optimization possibilities?

Start with the sort-based shuffle implementation and the performance-tool result identifying serialization as the largest CPU cost. Compare the current direct use of PrestoVectorSerde with possible alternatives, and define completion by a measured reduction in serialization overhead with equivalent shuffle behavior.

Written by the indexing model from the issue text.