Against All Odds: Overcoming Typology, Script, and Language Confusion in Multilingual Embedding Inversion Attacks