[VL] Scan output metric is not accurate.

### Backend

VL (Velox)

### Bug description

Sometimes, we will try to figure out the key contribution of the engine. The plan metrics is a good way to evaluate performance。
But we found that Scan metrics are inaccurate, leading to a series of misjudgments。

Q96 (1 TB scale) is one example.


Here: "number of output rows: 709, 525". This value is equal to the output of the third join, which is incredible.

cc @zhouyuan @weiting-chen , Thanks.

### Spark version

None

### Spark configurations

_No response_

### System information

_No response_

### Relevant logs

```bash

```

Reproduce Q96 at 1 TB scale and inspect the plan's scan metrics against the actual scan and join outputs shown in the report. Trace where the scan output-row value is produced; done means the scan metrics report the scan's own output accurately rather than a later join's output.

Written by the indexing model from the issue text.