Consider using candle built-in interpolate op for image resize

Hi, i found there were plenty of apps and models inside candle have image resize, for example:
yolo, sam, qwenvl, paddlevl-ocr etc, almost any image based models have resize.

However, current using image crate which is wide kwown slow.

This is problematic when the models runs in REALTIME, it become a massive bottleneck to run vision apps.

Would consider using built-in interpolate op as resize and using kornia-rs io as image decoding and reading