[PHI][Build] isolate oneDNN headers from CUDA translation units (#79659)

* docs: design CUDA oneDNN header isolation

* docs: plan CUDA oneDNN header isolation

* refactor: isolate oneDNN tensor descriptors

* refactor: remove oneDNN from common headers

* chore: exclude internal planning docs

* fix: include oneDNN context explicitly

* fix: include oneDNN context in reduce kernels

* fix: include oneDNN dependencies in layout test

* fix: preserve oneDNN descriptor access before allocation

* test: cover oneDNN descriptor helper boundaries

* fix: export oneDNN storage type info on Windows

* test: keep oneDNN rejection case backend neutral

---------

Co-authored-by: Nyakku Shigure <sigure.qaq@gmail.com>