Shuffle cannot be removed expectly.

### Backend

VL (Velox)

### Bug description

When testing tpcds, we found that the disk space utilization rate was increasing, which eventually led to the failure of tpcds due to insufficient disk space. After investigation, we found that this #6022 caused the shuffle to not be cleaned up in time, resulting in the continuous increase of block space.

BTW, we are running with Spark3.5.1 and I haven't test other Spark Versions.

### Spark version

None

### Spark configurations

_No response_

### System information

_No response_

### Relevant logs

_No response_

Start with issue #6022 and reproduce the TPC-DS workload using Spark 3.5.1 with the Velox backend, watching shuffle block disk usage. Done means shuffle data is cleaned up as expected and the workload no longer exhausts disk space.

Written by the indexing model from the issue text.