Suggestion: Add roofline/arithmetic-intensity analysis to matrixMul samples

The matrixMul and matrixMulCUBLAS samples report GFLOPS and execution time, and matrixMulCUBLAS already compares a custom kernel against cuBLAS — both genuinely useful for learners.

Suggestion: extend these (or add a companion sample) that additionally:

1. Implements the same GEMM at multiple optimization stages — naive, shared-memory tiled, and Tensor Core (WMMA) — so learners see the actual performance delta between each technique, not just a single implementation.
2. Reports arithmetic intensity (FLOPs/byte) alongside achieved GFLOPS, plotted against the GPU's roofline (peak compute vs peak memory bandwidth), so learners can see whether their kernel is memory-bound or compute-bound.

I built a small reference implementation exploring this for my own learning, benchmarked on an RTX 3050: https://github.com/ragulk143/cuda-gemm

Happy to contribute a PR along these lines if this is something the maintainers would find valuable for the samples repo.

Start by reading the matrixMul and matrixMulCUBLAS samples, then compare them with the linked cuda-gemm reference implementation. Define how the naive, shared-memory tiled, and WMMA stages will be presented and how arithmetic intensity and the GPU roofline will be reported; done means learners can compare each stage's performance and bound.