feat: cache-blocked SIMD MatMul achieving 14-17 GFLOPS

Add SimdMatMul.cs with GEBP (General Block Panel) algorithm:
- Cache blocking: MC=64, KC=256 tuned for L1/L2 cache
- 8x16 micro-kernel with 16 vector accumulators
- K-loop unrolled by 4 for better ILP
- FMA support when available
- Aligned memory allocation for packing buffers

Performance improvement (single-threaded):
- 256x256:  4.6 → 16.8 GFLOPS (3.7x)
- 512x512:  5.8 → 16.8 GFLOPS (2.9x)
- 1024x1024: 5.5 → 16.1 GFLOPS (2.9x)
- 2048x2048: 3.5 → 14.7 GFLOPS (4.2x)

This approaches OpenBLAS single-thread performance (~20-40 GFLOPS)
without requiring parallelization.

Author: Nucs

src/NumSharp.Core/Backends/Default/Math/BLAS/Default.MatMul.2D2D.cs

@@ -81,14 +81,12 @@ private static unsafe bool TryMatMulSimd(NDArray left, NDArray right, NDArray re
             {
                 case NPTypeCode.Single:
                 {
-                    var kernel = ILKernelGenerator.GetMatMulKernel<float>();
-                    if (kernel == null) return false;
-
                     float* a = (float*)left.Address;
                     float* b = (float*)right.Address;
                     float* c = (float*)result.Address;
 
-                    kernel(a, b, c, M, N, K);
+                    // Use cache-blocked implementation for better performance
+                    SimdMatMul.MatMulFloat(a, b, c, M, N, K);
                     return true;
                 }