fix(sortformer): honor packed streaming state

[ivan-digital]

Summary

• start Sortformer's valid speaker-cache prefix at zero and pass its real length to the ONNX graph
• read the export's packed prediction prefix instead of fixed cache/FIFO capacity offsets
• refresh pre-compression cache predictions to match NeMo's reference update
• update regression coverage and wrapper documentation

Evidence

• clean default build: 31/31 CTest tests pass
• ONNX targets build; Sortformer cache test passes
• 12-file Azerbaijani benchmark: corrected ONNX path 26.30% DER / 100% count agreement versus deployed API 57.20% / 83.3%; direct NeMo v2.1 is 27.06% / 100%
• on a representative recording, the corrected path removes the label permutation at the first 27.2-second boundary and matches NeMo's timeline with zero speaker confusion