Efficiently Train ASR Models that Memorize Less and Perform Better with Per-core Clipping