Feature request: CPU separable 3D linear resampling for DHWC volumes

## Request

Please consider a CPU resize or resample primitive for NumPy arrays with shape `(D, H, W, C)`. The operation should resize only the spatial axes, preserve the explicit channel axis, support `uint8` and `float32`, arbitrary channel counts, and unit-length output axes.

## Use case

Albucore performs CPU volume augmentation during model training. NumKong 7.7.0 exposes no public resize, interpolate, or resample operation, so Albucore compares pure NumPy, OpenCV packing, and Torch CPU paths instead.

## Performance context

For `64x64x80x9 -> 128x96x120x9`, `D*C=576` exceeds OpenCV's encoded channel limit. The current fallback is per-slice 2D OpenCV plus a depth pass. A NumKong separable 3D kernel could process the volume directly and avoid packing and Python slice-loop overhead.

## Useful initial scope

Half-pixel linear interpolation on CPU with float32 accumulation would cover the immediate use case. A documented uint8 final rounding rule would allow image volumes. Nearest interpolation, batch layouts, antialiasing, and GPU execution can be separate decisions.

Start from the requested DHWC shape, half-pixel linear interpolation, float32 accumulation, uint8 support, and unit-length output axes. Review NumKong's existing public CPU primitives and test structure before deciding the API and kernel boundaries; done means direct 3D resampling meets these requirements without packing or Python slice loops.

Written by the indexing model from the issue text.