Make OpenCV a core dependency, encode RGBA PNG, bump to 0.4.0

[MilkClouds]

Summary

• OpenCV is a core dependency. The images extra is gone (no alias); opencv-python-headless>=4.12 is in dependencies. cv2 is still imported lazily on first image-codec use; _opencv.py drops the "install tensorcodec[images]" message and keeps only the 4.12 version check.
• NumPy floor is now 2. Every opencv-python-headless 4.12+ wheel requires numpy>=2 on Python 3.9+, so numpy>=1.26 could no longer be satisfied. tensorcodec now declares numpy>=2, and the glibc 2.17 release check pins cp310 to NumPy 2.0.2 (the oldest NumPy 2 with cp310 manylinux2014 wheels) instead of 1.26.4. tensorcodec-av alone keeps numpy>=1.26.
• RGBA PNG encoding. PngEncoder accepts 4-channel CHW uint8 and converts RGBA→BGRA for OpenCV, so alpha is kept. JpegEncoder still allows only 1 or 3 channels; with 4 it raises ValueError: JpegEncoder needs a nonempty CHW uint8 array with 1/3 channels.
• Encode/decode symmetry. No decoder change was needed. decode_image(..., mode="RGBA") (or "UNCHANGED") returns 4×H×W RGBA with straight alpha last, which is exactly what PngEncoder takes. The default mode="RGB" drops alpha without premultiplying.
• Version 0.4.0 in pyproject.toml (project and tensorcodec-av pin), av/Cargo.toml, av/Cargo.lock, src/tensorcodec/__init__.py, plus docs/releasing.md and docs/system_ffmpeg.md. Not tagged or published. The repo has no changelog.
• Docs (README, docs/images.md, docs/package_size.md) and CI (build-python.yml installs plain tensorcodec; build-av.yml no longer installs OpenCV separately next to the wheels) are updated to match.

Tests

• New test_png_rgba_round_trip: a random RGBA image with fully transparent pixels. It is read back exactly by Pillow, and by decode_image in RGBA/UNCHANGED mode. Default RGB mode returns the first three channels unchanged.
• test_jpeg_rejects_alpha. Input-error cases now use 2 and 5 channels for PngEncoder.
• Locally (Python 3.12, OpenCV 5.0, NumPy 2.5, FFmpeg 7.1.1 fixtures): ruff check and format pass, and the image and version tests pass. The full suite (without --compare) passes when run against the published tensorcodec-av 0.3.0 wheel. To do that, __version__ was temporarily set back to 0.3.0, because tensorcodec-av 0.4.0 does not exist yet and this PR does not touch av/ sources.

Notes

• Breaking for users who are on NumPy 1.x, or who use a different cv2 variant (opencv-python, contrib). The pip metadata now pulls in the headless wheel, and the docs warn against mixing variants.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YAijSnAE4aAjuS1EcCTwqS