GPU Acceleration for kornia-tensor and kornia-imgproc

kornia-rs currently supports only CPU-based computation for image processing operations like resize, warp, and color transforms. This project implements a GPU backend for kornia-tensor and kornia-imgproc, motivated by a concrete Bubbaloop use case: a real-time camera pipeline that needs faster preprocessing to make VLM inference feasible on edge hardware. The work includes a feature-gated GpuAllocator with explicit host-device transfer semantics, allocator-based backend dispatch for tensor operations, and GPU kernels for resize, warp_affine, warp_perspective, and RGB <-> Grayscale conversion using CubeCL.