Fix pre-Blackwell validation in the Ada FP8 GEMM example

[Lec16sf]

Description

Restore the pre-Blackwell device and CUDA Toolkit validation in example 58.

Problem

Starting with CUTLASS v4.0.0, the props.major < 10 branch in examples/58_ada_fp8_gemm/ada_fp8_gemm.cu is empty:

bool satisfied;
if (props.major < 10) {
}
else {
  satisfied = (__CUDACC_VER_MAJOR__ > 12) ||
              (__CUDACC_VER_MAJOR__ == 12 &&
               __CUDACC_VER_MINOR__ >= 8);
}

As a result, satisfied may be read without being initialized when running on a pre-Blackwell GPU. This is undefined behavior.
A supported GPU with compute capability 8.9 or greater and CUDA Toolkit 12.4 or later may therefore incorrectly enter the unsupported-device path and print:
CUTLASS's FP8 SM89 example requires an NVIDIA GPU with compute capability 8.9 or greater and CUDA toolkit version 12.4 or later
The example then exits without running the GEMM.

Changes

Restore the pre-v4.0.0 validation logic for pre-Blackwell devices.

• Initialize satisfied for pre-Blackwell devices.
• Require CUDA Toolkit 12.4 or later for the pre-Blackwell path.
• Require compute capability 8.9 or greater for the pre-Blackwell path.
• Preserve the existing CUDA Toolkit 12.8 requirement for Blackwell and newer devices.

Testing

Built the example for SM89 with CUDA Toolkit 12.4 and ran it with reference checking enabled:

./build-sm89/examples/58_ada_fp8_gemm/58_ada_fp8_gemm \
  --reference-check=1

The results are as follows:

Running GEMM with staged accumulation (OpMultiplyAdd)
=====================================================
Problem size: 1024x1024x1024
Runtime (ms): 0.0116736
GFLOPs/sec:   183961
Passed

Running GEMM with fast accumulation (OpMultiplyAddFastAccum)
============================================================
Problem size: 1024x1024x1024
Runtime (ms): 0.0110432
GFLOPs/sec:   194462
Passed