[BACKEND] Fp8E5M2_to_Bf16 should fallback to the slow path for sm < 90 (#6858)

Author: Jokeren

third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/ElementwiseOpToLLVM.cpp

@@ -428,8 +428,9 @@ struct FpToFpOpConversion
              Fp16_to_Fp8E5M2_RTNE(computeCapability >= 89)},
             {{F16TyID, F8E5M2TyID, RoundingMode::RTZ}, Fp16_to_Fp8E5M2_RTZ},
             // F8 -> BF16
+            // mul{.rnd}.bf16 and mul{.rnd}.bf16x2 requires sm_90 or higher.
             {{F8E5M2TyID, BF16TyID, undefRounding},
-             Fp8E5M2_to_Bf16(computeCapability >= 89)},
+             Fp8E5M2_to_Bf16(computeCapability >= 90)},
             {{F8E4M3TyID, BF16TyID, undefRounding},
              Fp8E4M3Nv_to_Bf16(computeCapability >= 89)},
             // BF16 -> F8