fix: improve model loading with meta tensor handling and CPU fallbacks

refactor: simplify model device handling using accelerate library

fix: resolve meta tensor issues by forcing CPU device mapping

Initial commit

initial commit