Generate a CUDA kernel that computes a FUSED GEMM epilogue in a single kernel: out = gelu(A @ B + bias), where A and B are N x N fp16 matrices, bias is a 1-D fp16 tensor of length N broadcast across the rows of (A @ B) (i.e. bias[j] is added to column j), and gelu is the tanh approximation. Fuse the matrix multiply, the bias add, and the GELU activation into one shared-memory-tiled kernel so the N x N intermediate is never written to global memory. Three inputs: A (N x N fp16), B (N x N fp16), bias (length-N fp16).
