fix(ci): add pagination to check-cuda-versions GitLab API calls

[smoparth]

Summary

• Add pagination support to the check-cuda-versions workflow's GitLab API calls
• Extract API fetching into a fetch_all_pages() function that loops through pages until exhausted
• Include safety limit (10 pages / 1000 items) with ::warning:: annotation if reached

Details

The workflow previously made a single curl call with per_page=100. While NVIDIA currently has only ~19 entries in their dist/ directory, this would silently truncate results if the count exceeded 100.

The new fetch_all_pages() function:

• Loops through pages incrementing page parameter
• Merges JSON arrays across pages using jq -s 'add'
• Exits on empty response, zero-length array, or fewer than per_page items
• Has a safety limit of 10 pages to prevent infinite loops
• Emits a GitHub Actions warning if the safety limit is reached

Both the primary (URL-encoded path) and fallback (numeric ID) API calls use the same function.

Test plan

• Tested fetch_all_pages() against live NVIDIA GitLab API — returns all 19 items
• Verified JSON output format is backward compatible with downstream processing
• Tested edge case: invalid project ID returns [], enabling fallback
• Validated YAML syntax

Fixes opendatahub-io#28

🤖 Generated with Claude Code