Test 2025.2

[rolandschulz]

We should test whether there are any perf regressions with 2025.2. In particular whether
https://github.com/codeplaysoftware/cutlass-sycl/blob/8f5fe44d47b98622a5841f0c39f14b08b12d2e2c/include/cutlass/numeric_conversion.h#L542
is actual fixed and not using the w/a is OK.