`torch.backends.mkldnn.flags()` CM should not warn (#150358)

By returning `None` rather than `False` from `THPModule_allowTF32OneDNN` when USE_XPU is not defined

Added regression test

Fixes https://github.com/pytorch/pytorch/issues/149829

Pull Request resolved: https://github.com/pytorch/pytorch/pull/150358
Approved by: https://github.com/atalman

(cherry picked from commit 6470b373c16017f5cb8f1aa4060bb60632b18160)

Co-authored-by: Nikita Shulga <nshulga@meta.com>