fix: cap preview_num_records to entity row count in replace workflow

[lipikaramaswamy]

Summary

When preview() is called with Substitute replacement, llm_replace_workflow.py filters the input down to only entity-bearing rows before passing to the adapter. Previously it passed the original preview_num_records (total dataset size) unchanged, causing Data Designer's seed generator to wrap around and produce duplicate rows.

The fix caps preview_num_records to len(entity_rows) before the adapter call.

Type of Change

• Bug fix
• New feature
• Breaking change
• Documentation update
• Refactoring

Testing

• Tests pass locally
• Added/updated tests for changes

The fix is a one-liner capping a value passed to a mocked dependency. A unit test would only assert that min() works, not that duplicates are actually prevented — the real invariant lives at the Data Designer seed wrap-around boundary which requires a live NIM endpoint to exercise.

Related Issues

Closes #58