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

Goal
- Fuse affine, cosine activation, sigmoid gate, and multiply to reduce memory traffic and kernel overhead.

Inputs/Outputs
- Input `x`: [B, D], float32
- Parameters `scale`, `bias`: [D], float32
- Scalars `alpha`, `beta`: float32
- Output `y`: [B, D], float32

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

CUDA Design
- 2D grid; block=128; ILP=1; float4 vectorization
- Use fast `__cosf`; affine via FMA

Validation
- Accuracy `torch.allclose(rtol=1e-3)`
- Speedup ≥ 1.30x at B=16, D=16384
 
 Extended Benchmark & Requirements
- Test 3 shapes (D=4096/16384/65536) and FP32/FP16/BF16 when supported
- Report per-case timing and speedup; 100 iterations, synchronized
- FP32 `rtol=1e-3`, FP16/BF16 `rtol=1e-2`
- If speedup <1.3x, emit bottleneck analysis and improvement suggestions
