Change DataFileSet from std::unordered_set to a custom class that preserves
insertion order, similar to Java's DataFileSet which uses LinkedHashSet.
This is important for row ID assignment in v3 manifests, where row IDs
are assigned based on the order files are written.

The implementation uses both a vector (for insertion order) and an
unordered_set (for O(1) duplicate detection) to maintain the same API
while preserving order.