Running the regression test suite create 150k files - reduce and/or auto clean up

### Bug description

After running the whole regression test suite:
```
$ find regress-20250616-aiJ | wc
148457 148457 13959112
```

On any system, that's a lot of files. On Trixie, that's 15% of our inode quota on work, with just one regression run.

Two requests:
- [ ] Clean up automatically as part of the regression suite
- [ ] Change the everyvoice preprocessing pipeline to keep fewer files on disk, e.g., using a dataloader instead.

Start by locating the regression test suite and the EveryVoice preprocessing pipeline, then run the suite and measure the generated files. Done means the regression run automatically cleans up its artifacts and the preprocessing pipeline keeps substantially fewer files on disk, with tests or documented measurements confirming both changes.

Written by the indexing model from the issue text.