tensor upsample_cubic support

hi, would consider support upsample_cubic for tensor? this is very needed as many resize op using cubiclinear, currently the resize had to using image which is slow, but upsample doesn't support cubic at the moment.

Start by locating Candle’s tensor upsample implementation and the existing image-based resize path. Compare how current interpolation modes are represented and identify the tests covering upsampling. Done means tensor upsampling supports cubic interpolation without requiring the image path, with coverage for the new behavior.

Written by the indexing model from the issue text.