Tracking: N-D numerical, image, and volume kernels requested by AlbumentationsX

## Goal

Track the NumKong primitives that would let AlbumentationsX and Albucore process images, batches, videos, and volumes without Python loops or chains of NumPy temporaries.

The common layouts are channel-last:

- Image: `(H, W, C)`.
- Image batch or video: `(N, H, W, C)`.
- Volume: `(D, H, W, C)`.

## Numerical Tensor layer

- #368 — buffer-first N-D cast with `out=` and explicit conversion semantics.
- #311 — axis-aware `moments` and `minmax`.
- #315 — mixed-dtype broadcasted affine transform for normalization.
- #327 — shape-preserving N-D `scale` and explicit integer rounding.
- #366 — broadcasted comparisons, `minimum`, `maximum`, `clip`, `where`, and fused compare-select.
- #369 — N-D `lerp` with scalar, channel, image, batch, or volume weight tensors.
- #365 — elementary math including `abs`, `sqrt`, `exp`, `log`, `floor`, `ceil`, `rint`, `atan2`, and `hypot`.
- #363 — histogram and `bincount`.
- #247 — N-D `take`/gather for channel permutations.
- #312 — direct `matmul` for one-shot and changing matrices.
- #299 — stateful SIMD random generators.

## Spatial and neighborhood layer

- #362 — batched 2D and true 3D resampling for resize, affine/perspective warp, and dense remap.
- #364 — N-D `convolve1d` and separable box/Gaussian filters.
- #367 — N-D minimum, maximum, median, and mode neighborhood filters; morphology derives from min/max filters.

## Current 3D dependency chain

AlbumentationsX is expanding from slice-wise volume handling to true 3D augmentation. For that work, #362 and #364 are the two foundational kernels rather than general future optimizations:

- #362 unlocks `Resize3D`, `Affine3D`, `ElasticTransform3D`, and `Anisotropy3D`.
- #364 unlocks true `GaussianBlur` over D/H/W and smoothing for 3D elastic displacement fields.
- #367 later unlocks true 3D morphology and median/mode neighborhood transforms.

Equivalent OpenCV requests are tracked in [opencv/opencv#29605](https://github.com/opencv/opencv/issues/29605) and [opencv/opencv#29604](https://github.com/opencv/opencv/issues/29604). Having both requests lets Albucore benchmark semantics, runtime, memory, and release availability before choosing a backend; it does not require NumKong to copy OpenCV's API.

## Related backend work

- [StringZilla #302](https://github.com/ashvardanian/StringZilla/issues/302) — typed and per-channel LUT, including `uint8 -> float32`. StringZilla already owns the public LUT/`translate` path used by Albucore.

## Common API contract

These issues can land independently. The Python-facing APIs benefit from the same rules:

- Preserve arbitrary leading batch dimensions.
- Make spatial or reduction axes explicit.
- Support C=1, channel counts greater than four, and C>128 without reinterpreting the Tensor rank or shape.
- Accept NumPy buffers directly.
- Offer `out=` where aliasing is well-defined.
- Document dtype promotion, accumulation dtype, rounding, saturation, NaN behavior, and border behavior.
- Release the GIL around the complete native operation.
- Start with contiguous input when needed and state stride restrictions explicitly.
- Keep the native layer free of hidden thread ownership; callers can partition work when an API exposes useful row or batch ranges.

## Suggested implementation order for our workloads

1. Buffer-first cast, axis-aware statistics, mixed-dtype affine, comparisons/clamp/select, and RNG address the largest reusable numerical passes. StringZilla #302 covers the typed LUT path.
2. Resampling and separable filtering form the minimum native foundation for the new true 3D transform family.
3. Dense-weight `lerp`, direct matmul, histogram/bincount, channel gather, and elementary math cover frequent specialized passes.
4. Neighborhood filters, including mode filtering and true 3D morphology, extend the spatial family after the resampling/filtering foundation.

This issue is an index. Each linked issue contains its own minimum useful contract, correctness cases, and benchmark baselines.

This issue is an index rather than a self-contained implementation task. Read the linked issues, especially #362 and #364, for the actual contracts, correctness cases, and benchmarks; completion means delivering one of those independently scoped kernels rather than changing this tracker.

Written by the indexing model from the issue text.