CUDA Sigmoid: Implement Optimized Sigmoid Kernels for Float32 and Float16

[debashishc]

Implement the following optimized sigmoid kernels for float32 and float16 with vectorized versions and PyTorch bindings for improved performance.

• sigmoid_f32_kernel: Standard sigmoid function for float32 data type.
• sigmoid_f32x4_kernel: Vectorized sigmoid for float32, processing 4 elements at a time (float4).
• sigmoid_f16_kernel: Standard sigmoid function for float16 (half-precision).
• sigmoid_f16x2_kernel: Vectorized sigmoid for float16, processing 2 elements at a time (half2).
• sigmoid_f16x8_kernel: Unpacked version of float16, processing 8 elements in parallel.
• sigmoid_f16x8_pack_kernel: Packed version of sigmoid_f16x8_kernel for efficient memory access.
• PyTorch bindings: Expose the above kernels through PyTorch.