Track KumoTabular downloads through model configuration

[aw471]

KumoTabular currently requests only size-specific classifier.pt or regressor.pt files from Hugging Face. Enable the shared checkpoint downloader's config.json request before uncached KumoTabular downloads, using the same repository, revision, and cache directory. Cached checkpoint loads remain offline.

Stacked on kumorelational-download-tracking, which introduces the shared configuration download support and cache behavior coverage. This PR adds only KumoTabular's opt-in.

Draft dependency: nvidia/Kumo-Tabular must publish root-level config.json in a new model release, and this PR's current v1.0.0 pin must be updated to that release before merging. The current v1.0.0 release does not contain config.json, so uncached loading with this draft would fail today. Do not merge until both steps are complete.

The metric counts requests to config.json, not completed weight transfers or unique users. Hugging Face's counting rules: https://huggingface.co/docs/hub/models-download-stats.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.