Objective: Fused LayerNorm + SiLU CUDA kernel with rtol=1e-3 accuracy and ≥1.3x speedup.

Computation:
- Row layer normalization (mean/var over dim) with epsilon, then affine transform, then SiLU: `y = u * sigmoid(u)`.

Numerical Method:
- Use Welford’s online algorithm for mean/variance to match PyTorch precision and avoid catastrophic cancellation.
- Perform warp-level reductions and shared-memory aggregation across warps.

Performance:
- 512 threads per block; coalesced loads and parallel write-back.
- Fuse normalization, affine, and activation to minimize memory traffic.

Benchmark:
- Batch size: 16; Dim: 16384; 100 iterations; target speedup ≥1.3x.
