Add optional OpenCV image decoding and JPEG/PNG encoding

[MilkClouds]

Image support reuses an installed OpenCV build without adding native image builds, vendored sources or a second codec backend. This replaces #12; native video/audio code is unchanged.

• Add CPU decode_image/jpeg/png/webp/gif/avif, NumPy CHW/NCHW output, orientation and animation handling, plus JpegEncoder and PngEncoder.
• Import OpenCV lazily; reuse cv2 >= 4.13 or install the optional images extra. NumPy remains the only required Python dependency. Pillow is test-only.
• HEIC is explicitly unsupported: no ctypes/libheif binding, system-library setup or binary fixtures. APNG, high-bit-depth AVIF, CMYK JPEG UNCHANGED and nondefault per-call AVIF thread settings also raise explicit errors. AVIF alpha removal follows OpenCV's straight-RGB behavior, documented as a difference from TorchCodec.
• Keep input/mode validation, color conversion and orientation handling required to adapt OpenCV to the public API. Image runtime implementation is approximately 544 Python lines, plus decoder exports. Include source-generated tests and a 65-line corpus benchmark.

Validation: Ruff lint/format passed; full local pytest --compare: 444 passed, including a regression for 1/2/4-bit grayscale PNG transparency. Tests cover independent encoder output decoding, color/depth/layout, all-frame animation, orientation, failures and lazy optional imports.

The final adapter matched TorchCodec 0.17.0 exactly for RGB output on 54 JPEG/PNG/WebP inputs. Single-CPU geometric-mean latency ratios vs TorchCodec: JPEG 1.02–1.05×, PNG 1.28×, WebP 1.07–1.11×. These are corpus-specific results; encoding speed was not benchmarked. Details and limitations: docs/images.md.