Fix: Correct Bicubic Boundary Interpolation in CV-CUDA Resize (Tensor Path)

[sezer-muhammed]

Summary

This PR for the bicubic boundary interpolation fix for CV-CUDA’s Resize operator.
The tensor path now performs per-tap clamping on all 16 cubic samples, matching OpenCV’s behavior at image edges and eliminating the flattening artifacts seen in legacy kernels.

Image Comparison (Original → OpenCV → CvCuda → New CvCuda)

---

Dataset	Metric	OpenCV vs Old	OpenCV vs New
real_tabby	Avg abs diff	0.020	0.002
synth	Avg abs diff	1.696	0.116

The new implementation reduces mean absolute error (compared to OpenCV) by an order of magnitude compared to the legacy kernel, spesifically at the images edges.

---

Note: I used OpenCV as colse to ground truth source for comparisons.