Write resized video frames into final output

[bghira]

Summary

• allocate each resized video result once
• point OpenCV destination Mat headers directly at each final output frame
• eliminate temporary resized frame ndarrays and reconstruction copies
• preserve support for videos with a strided frame axis when each frame remains contiguous

Root cause

The video path flattened all frames, used the copy-heavy image API to create one ndarray per resized frame, then allocated each output video and copied every frame into it again. The memory traffic and allocation overhead dominated resize time.

Performance

Median local latency on Apple M3 Max:

• 30x1024x1024 -> 30x512x512: 11.50 ms -> 1.57 ms
• 120x512x512 -> 120x224x224: 7.72 ms -> 2.48 ms

The direct path also beat resizing through batch_resize_images_zero_copy followed by numpy.stack (3.37 ms and 3.07 ms respectively).

Validation

• cargo test --features opencv,simd (60 passed)
• Python test suite (68 passed)
• repository pre-commit hooks, including competitive performance benchmarks

[Copilot]

Pull request overview

This PR optimizes the Rust/OpenCV video resize path by allocating each output video once and having OpenCV write resized frames directly into the final output buffer, reducing per-frame allocations and extra copies.

Changes:

• Replace the “flatten all frames then reconstruct” approach with per-video output allocation and direct frame writes.
• Create OpenCV Mat headers that point directly into the destination ndarray::Array4 frame buffers.
• Keep support for strided frame-axis inputs as long as each individual frame is contiguous.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.