Fix Filtering Logic for Normalized WER Calculation

[mahdibaghbanzadeh]

This PR fixes a bug in the filtering step for computing normalized WER. Previously, predictions and references were filtered separately, which could cause misalignment. Now, both lists are filtered together to ensure they stay in sync.

Changes:

• Filter predictions and references together using a single list comprehension.
• Prevent index mismatches by unpacking the filtered pairs correctly.
• Handle edge cases where all references are empty.

Why this is needed:
This ensures that only valid reference-prediction pairs are evaluated without breaking alignment.