Fix symlink containment in UbuntuCorpusTrainer.extract()

[gunthercox]

No description provided.

[arpitjain099]

Ran the race against the updated branch. It holds.

Same harness as before, swapping data_directory for a symlink once extraction has visibly started, 30 runs at each size:

members=   50  attacker won 0/30 runs, 0 files written outside
members=  500  attacker won 0/30 runs, 0 files written outside
members= 4000  attacker won 0/30 runs, 0 files written outside

Against 24/30, 27/30 and 27/30 before, with 1168, 13472 and 107971 files landing outside. The swap itself still landed on all 90 runs, so that is the anchoring doing the work rather than the harness missing its window.

I also forced the flag off to exercise the non-dir_fd path, since that is what macOS and the BSDs take: 0/30 at every size there as well. So the staging directory plus the atomic rename closes it on its own, and the fd anchoring is defence in depth rather than the load-bearing part. Worth knowing if you ever need to simplify.

One thing to fix before this merges, which I opened as #2451: _DIR_FD_EXTRACTION_SUPPORTED is assigned twice and the second assignment drops the /proc/self/fd term, so the flag comes out true on macOS where procfs does not exist. The mkdir succeeds because it is genuinely fd-relative, then tarfile writes through /proc/self/fd/N and extract() dies on the first write with OSError: [Errno 30] Read-only file system: '/proc'.