Fix Webshart video sample and cache handling

[bghira]

Summary

This PR fixes several Webshart data-path, cache, metadata, and distributed preparation edge cases that were found while separating the H3 branch into independently reviewable work.

• Normalize path-normalized webshart:/ sample IDs back to canonical webshart:// identifiers before cache key and backend lookup logic runs.
• Preserve webshart_shard_cache_gb=0 as an explicit request to disable whole-shard caching, while rejecting negative cache sizes.
• Add video-aware sample loading and metadata handling for Webshart shards, including ffprobe-backed bucket metadata when indexed metadata is incomplete.
• Add sparse HTTP range probing for remote TAR members when offset and length metadata are available, so ffprobe can inspect container metadata without downloading the full video sample.
• Prepare Webshart video bucket metadata in bounded parallel chunks using the backend download parallelism setting.
• Make max-sample limiting and shuffled distributed bucket splitting independent of discovery/cache reload order so ranks partition the same serialized cache contents.
• Reload refreshed bucket caches before splitting and avoid splitting rank-local prompt records a second time during text embed cache preparation.
• Reset compressed torch cache streams before fallback loading so cache reads do not depend on the caller's stream position.
• Document the zero shard-cache behavior.

Validation

Ran on the PR branch:

.venv/bin/python -m unittest -v -f tests.test_webshart_backend tests.test_backend_builders tests.test_factory_edge_cases tests.test_metadata_backend tests.test_text_embeds tests.test_utils_pathing

Result: 206 tests passed.

Also checked the committed files for /Users/kash and /Users/ path leaks; none were found.