v0.0.27 - LiteRT wasm multithreading for up to 2x faster CPU inference (#300)

🌟 Summary

Ultralytics Inference v0.0.27 focuses on faster browser CPU inference with optional LiteRT WebAssembly multithreading, plus stronger CUDA reliability, cleaner postprocessing, better docs, and release/CI hardening 🚀

📊 Key Changes

• LiteRT WASM multithreading for browser inference 🧵
  The @ultralytics/yolo web package now passes an optional threads setting to LiteRT.js and automatically enables WebAssembly multithreading when the page is crossOriginIsolated.

• Version bump to 0.0.27 across Rust and npm packages 📦
  Updated ultralytics-inference, ultralytics-inference-web, and @ultralytics/yolo package versions, along with README and CUDA documentation examples.

• CUDA and TensorRT inference reliability improvements ⚡
  CUDA/TensorRT compute-stream binding was refactored into shared logic, preserving fail-fast behavior when CUDA preprocessing is active to avoid silent CPU fallback issues.

• New CUDA-focused tests 🧪
  Added GPU tests for CUDA preprocessing, model loading, inference, and CUDA/CPU parity with YOLO26 ONNX workflows.

• Postprocessing cleanup for boxes and keypoints 🎯
  Detection, segmentation, pose, and YOLO26 end-to-end output handling now reuse shared scaling/clipping helpers for more consistent results and easier maintenance.

• Annotation crash prevention 🛡️
  Label drawing now safely skips empty or invalid text dimensions, preventing runtime panics caused by zero-width or zero-height annotation labels.

• YOLO26 preprocessing/postprocessing optimizations 🦀
  Several YOLO26 end-to-end shape helpers are now compile-time friendly, and RGB pixel iteration is more explicit and type-safe.

• Improved build and usage documentation 📚
  Added clearer source-build prerequisites for Linux, macOS, and Windows, updated device option docs to include rocm:0 and xnnpack, and refreshed CUDA guidance.

• Web classification docs corrected 🌐
  The web README now documents top5names in classification probability outputs.

• Release and CI workflow hardening 🔒
  Releases are now restricted to the main branch, and GitHub Actions dependencies were updated with pinned versions for more reproducible automation.

🎯 Purpose & Impact

• Faster browser CPU inference where supported 🚀
  Browser users running LiteRT-backed .tflite models may see significantly better CPU performance, potentially up to about 2x faster in environments that support WebAssembly threads and SharedArrayBuffer.

• No extra setup for supported browsers ✅
  Multithreading is enabled automatically when the web page is properly cross-origin isolated. If not supported, LiteRT continues to run single-threaded, keeping compatibility intact.

• More dependable GPU inference ⚡
  CUDA and TensorRT users benefit from better-tested GPU paths and safer stream handling, reducing the chance of confusing fallback behavior or crashes.

• More stable visual outputs 🖼️
  Annotation rendering is more robust, especially for edge cases involving empty labels or problematic fonts.

• More consistent YOLO26 results handling 🎯
  Shared scaling and clipping logic reduces duplicated code and helps keep detection boxes and pose keypoints consistent across model output formats.

• Easier installation from source 🛠️
  Clearer system dependency instructions should help new users avoid common Rust build failures on Linux, macOS, and Windows.

• Safer, more predictable releases 🔐
  Restricting release automation to main helps ensure published crates, npm packages, and GitHub tags come from the official release branch.

What's Changed

• Bump taiki-e/install-action from v2.82.6 to v2.82.7 in /.github/workflows by @UltralyticsAssistant in #279
• Improve coverage accounting for runtime paths by @glenn-jocher in #280
• docs: readme clean up by @onuralpszr in #281
• build(deps): update package versions in Cargo.lock by @onuralpszr in #283
• refactor(postprocess): dedup box and keypoint helpers by @onuralpszr in #282
• fix(annotate): guard label drawing against zero-width rect by @onuralpszr in #284
• Add AGENTS.md agent guidance and fix documentation drift by @glenn-jocher in #287
• Bump taiki-e/install-action from v2.82.7 to v2.82.8 in /.github/workflows by @UltralyticsAssistant in #286
• Restrict release workflow to the main branch by @glenn-jocher in #288
• docs(readme): document build prerequisites for macOS, Linux, and Windows by @onuralpszr in #289
• Bump astral-sh/setup-uv from v8.2.0 to v8.3.0 in /.github/workflows by @UltralyticsAssistant in #290
• build(deps): Bump taiki-e/install-action from 2.82.8 to 2.82.9 in /.github/workflows by @dependabot[bot] in #291
• chore: update dependencies in Cargo.lock and Cargo.toml by @onuralpszr in #292
• Use const fn and as_chunks in pre/postprocessing by @onuralpszr in #296
• refactor CUDA/TensorRT EP stream binding by @onuralpszr in #295
• Add CUDA inference tests by @onuralpszr in #297
• Bump taiki-e/install-action from v2.82.9 to v2.82.10 in /.github/workflows by @UltralyticsAssistant in #298
• Bump astral-sh/setup-uv from v8.3.0 to v8.3.1 in /.github/workflows by @UltralyticsAssistant in #299
• LiteRT wasm multithreading for up to 2x faster CPU inference by @onuralpszr in #300

Full Changelog: v0.0.26...v0.0.27