Almost 10 times Hash Agg spill memory than vanilla Spark

### Backend

VL (Velox)

### Bug description

we found that for some case, Hash Agg hash huge memory spilled, and the stage with huge memory spill is running slower than vanilla Spark

### Gluten version

_No response_

### Spark version

None

### Spark configurations

_No response_

### System information

_No response_

### Relevant logs

```bash

```

Start by obtaining a reproducible Velox hash-aggregation case and the missing Gluten, Spark, configuration, system, and log details. Compare its spill memory and stage runtime with vanilla Spark; done means the excess spill and slowdown are reproduced and their cause is identified.

Written by the indexing model from the issue text.