XPU uses the same Triton codegen path as CUDA, which asserts on direct
fp8-to-fp8 type conversions (float8_e4m3fn <-> float8_e5m2) in
_get_min_elements_per_thread. However, test_bad_cast only expected
this error for CUDA devices. XPU fell into the else (CPU) branch,
which expects the conversion to succeed, causing the test to fail.
Fix by...