[Bug]:Remove redundant heap allocations in kornia-apriltag inner loops

🐛 Describe the bug

The kornia-apriltag crate currently suffers from performance bottlenecks due to rapid heap allocations inside its inner detection loops. There are several TODO: Avoid allocation comments in the hot paths of decoder.rs and quad.rs. Because these inner loops (like fit_quads, compute_line_fit_prefix_sums, and decode_tags) run potentially thousands of times per frame during tag extraction, the continuous allocation of Vec arrays forces the memory allocator/OS into heavy churn, severely fragmenting the heap and bottlenecking frames-per-second throughput.

🔄 Steps to Reproduce

1.Run the ApriltagDetector on any image.
2.Observe the internal execution flow: the program allocates and deallocates multiple vectors (like quads, lfps, and detections) repeatedly inside the geometry and decoding functions for every single candidate.

💻 Minimal Code Example

N/A - This is a structural memory issue rather than a code crash. The problematic lines in the repository are:

- crates/kornia-apriltag/src/decoder.rs(Line 354): let mut detections = Vec::new(); // TODO: Avoid allocations on every call

- crates/kornia-apriltag/src/quad.rs(Line 126): let mut quads = Vec::new(); // TODO: Avoid this allocation every time

- crates/kornia-apriltag/src/quad.rs (Line 414): let mut lfps = vec![LineFit::default(); gradient_infos.len()]; // TODO: Find a way to avoid allocation

✅ Expected behavior

These dynamic arrays should be hoisted out of the low-level functions and stored as persistent, reusable buffers inside the ApriltagDetector state (e.g., passing a &mut Vec buffer downward). During each detection frame, the pipeline should just call .clear() and reuse the existing heap capacity over and over without any new OS memory allocations.

❌ Actual behavior

The code calls Vec::new() and vec![...] directly inside the hot loops, causing gigabytes of garbage allocations over a sustained video stream.

🔧 Environment

- kornia-rs version:
- Rust version (`rustc -V`):rustc 1.93.1
- Cargo version (`cargo -V`):cargo 1.93.1
- OS (e.g., Linux, macOS, Windows):macOS
- Target architecture (if cross-compiling): N/A
- Python version (if using Python bindings): Python 3.9.6

📝 Additional context

This optimization targets purely memory footprint and single-thread scaling behavior, fulfilling the existing TODO markers left in the codebase by the original author. We will move those temporary vectors into the AprilTagDecoder struct itself. Instead of: Allocating new memory every frame We now: Allocate memory once at startup

Reuse the same buffers forever

No more repeated heap allocations. No more memory churn.

🤝 Contribution Intent

• I plan to submit a PR to fix this bug
• I'm reporting this bug but not planning to fix it