Add Fused Separable Convolution

🚀 The feature

Add a fused separable convolution op for fast Gaussian blur, and many other kernels.

ops.separable_conv2d(x, fx, fy, stride=1, padding=0)
ops.separable_conv_transposed2d(x, fx, fy, stride=1, padding=0)
ops.separable_conv3d(x, fx, fy, fz, stride=1, padding=0)
ops.separable_conv_transposed3d(x, fx, fy, fz, bias=b, stride=1, padding=0)

Motivation, pitch

Many filters can be implemented as separable filters. Gaussian, sinc, sobel, etc.

Right now the approach is to launch multiple consecutive F.conv2d, but this is often slower than a single full filter due to the overhead of kernel launch, memory writes, etc.

Separable convolutions can be written as a single CUDA kernel.

I believe there is already a function in CUDA, see: https://developer.download.nvidia.com/compute/cuda/1.1-Beta/x86_64_website/projects/convolutionSeparable/doc/convolutionSeparable.pdf

Alternatives

No response

Additional context

No response

The issue proposes new ops.separable_conv2d, ops.separable_conv_transposed2d, ops.separable_conv3d, and ops.separable_conv_transposed3d APIs, replacing multiple F.conv2d launches with fused kernels. Start by locating the existing F.conv2d path and reviewing the linked CUDA separable-convolution reference; the issue names no repository files or tests, so the implementation scope and validation plan need to be established first.