webshart caption lookup

[bghira]

This pull request introduces support for optimizing caption loading in Webshart-backed datasets, reducing costly range reads for datasets that store captions in sidecar files. It adds a new webshart_optimize_captions option, updates configuration handling, and ensures that sidecar captions are consolidated into the local metadata cache when enabled. The documentation is updated in multiple languages to explain the new feature and its CLI usage. Additional minor improvements include more robust tag frequency memoization and a small fix in aspect bucket computation.

Webshart caption optimization support:

• Added the webshart_optimize_captions option to dataset configs, CLI, and backend initialization, enabling automatic consolidation of sidecar captions into the local metadata cache for faster access. [1] [2] [3] [4] [5] [6] [7]
• Implemented _optimize_caption_metadata() in WebshartDataBackend to perform the coalescing step at startup, with appropriate checks for required Webshart build features.

Documentation updates:

• Updated the dataloader documentation in English, Spanish, Hindi, Japanese, Portuguese, and Chinese to describe the new caption optimization option, its benefits, and CLI usage for pre-coalescing captions. [1] [2] [3] [4] [5] [6]

Performance and bug fixes:

• Memoized the tag frequency computation in training save hooks to avoid expensive recomputation on remote datasets.
• Fixed a bug in aspect ratio bucket computation to properly check for caption sidecar files in Webshart metadata.