Operator: Mish-Affine-Gate (Fused CUDA Kernel)

Goal
- Fuse affine, Mish activation, sigmoid gate, and multiply to reduce bandwidth and launches; target ≥1.30x speedup.

Definition
- z = x * scale + bias
- m = z * tanh(softplus(z))
- g = sigmoid(alpha * m + beta)
- y = x * g
