test: Add test for ref_input parameter in fused linear preference

[xingyaoww]

This PR adds a test for the ref_input parameter that was introduced in #467.

Changes

• Add test_ref_input.py to verify the ref_input parameter works correctly in LigerFusedLinearPreferenceBase
• Test ensures that:
  • Policy model outputs (chosen_logps, rejected_logps) are identical when using ref_input vs not using it
  • Final loss and aux outputs differ when using ref_input vs using input_chunk for reference model
• Uses same parametrization as other tests for consistency

Testing

The test verifies that:

1. When ref_input is not provided, the reference model uses input_chunk
2. When ref_input is provided, the reference model uses it instead of input_chunk
3. The policy model outputs remain unchanged regardless of ref_input
4. The final loss differs when using different inputs for the reference model

Follows up on #467 which added ref_input parameter support.