[upstream #3549] [XPU] Enable functionalize_rng_ops on XPU (philox_rand prim + RNG decompositions)

[Stonepia]

[upstream pytorch#3549] [XPU] Enable functionalize_rng_ops on XPU (philox_rand prim + RNG decompositions)

Fixes https://github.com/intel-sandbox/torch-xpu-ops-exp/issues/384

Root Cause: The functionalize_rng_ops path hardcodes CUDA-only device checks via throw_on_non_cuda in three locations (decompositions_for_rng.py lines 34, 55; rng_prims.py line 118). XPU supports Philox RNG but is rejected by these gates. Additionally, philox_rand_offset likely queries CUDA-specific device properties.

---

Diff stat:

torch/_decomp/decompositions_for_rng.py | 11 ++++++++--
 torch/_prims/rng_prims.py               | 38 +++++++++++++++++++++++++++------
 torch/_prims_common/__init__.py         | 20 ++++++++++++++++-
 3 files changed, 59 insertions(+), 10 deletions(-)