Include archive path in WebDataset TAR read errors

[sankalpsthakur]

Fixes #7280

What changed

WebDataset catches tarfile.ReadError around TAR processing and re-raises with the resolved archive path; original remains on __cause__.

Covers failures while advancing the archive iterator and while reading a member payload (feature inference and example generation).

Uses str(tar_path) rather than repr(tar_path) so tracked_str origins (presigned URLs/credentials) are not exposed in tracebacks.

Before

ReadError: empty file

After

ReadError: Failed to read TAR archive "/path/to/corrupted-shard-00042.tar": empty file

Tests

• corrupt first shard (feature inference)
• valid shard then corrupt shard (example generation)
• truncated member during f.read()
• original ReadError preserved as cause
• Windows-safe path matching
• tracked_str origin with password/signature not exposed

Validation

• pytest (WebDataset / DownloadManager / StreamingDownloadManager): 46 passed, 10 optional skips
• make style / git diff --check: passed

[sankalpsthakur]

CI waiting on maintainer workflow approval (action_required). Local WebDataset ReadError-path tests are green.

[sankalpsthakur]

related: this covers the WebDataset TAR ReadError path from #7280; broader _iter_from_urlpath remainder left for a follow-up.

[lhoestq]

I would still include the origin, in most cases it's a hf:// URLs that points to the file online which is useful for debugging

[sankalpsthakur]

Updated — errors now include both the archive path and the hf:// origin.

[lhoestq]

feel free to do it in tracked_str.__repr__ directly insetad of having a custom logic just for webdataset, this way it will benefit other loaders as well

[sankalpsthakur]

Good idea — I'll move this into tracked_str.__repr__ in a follow-up so other loaders benefit too.

[sankalpsthakur]

Went ahead and implemented this now rather than deferring to a follow-up (061468a): the origin sanitization now lives in tracked_str.__repr__ (via a small _sanitize_origin helper), so any loader that reprs a tracked path gets the safe origin — userinfo and query strings are stripped while scheme/host/path (e.g. hf://…) are kept for debugging. WebDataset's except tarfile.ReadError block is back to a one-liner and its local sanitizer is gone. The existing WebDataset read-error tests cover the behaviour unchanged.