ciflow/xpu/186891: xpu: gracefully handle missing GenerateIntelXe in CUTLASS

When the CUTLASS library doesn't have GenerateIntelXe (e.g. NVIDIA CUTLASS

instead of Intel Xe CUTLASS), _gen_ops_cached raises NotImplementedError,

causing tests that force CUTLASS backend (like cpp_wrapper + max_autotune)

to fail on XPU with 'Arch 12 is not supported'.

Fix: Return empty ops dict {} with a warning instead of raising, so the

lowering gracefully falls back to Triton templates.

Co-authored-by: Copilot [email protected]