trunk/f62743b24c96df647c620aac8225a3a8b736a162: Fix XPU tolerance for large batch MHA compile test (#183722)

The disabled XPU test failure is a small fp16 compile/eager numerical delta, not the CUDA/cuDNN backend failure from the abandoned PR. Keep the existing strict comparison on other devices while using a narrow XPU tolerance for the output and gradient checks.

Fixes

Approved by: https://github.com/aorenste