Add ChromaSubsamplingMethod option for box-average downsampling

[tylerneylon]

Adds an opt-in ChromaSubsamplingMethod::Average that box-averages each chroma block when subsampling, matching libjpeg / libjpeg-turbo's default downsampler. The existing point-sampling behavior stays the default (Nearest), so there is no change for current users.

Motivation

When chroma is subsampled (e.g. 4:2:0), each output Cb/Cr sample represents a 2×2 block of source pixels. Today jpeg-encoder takes the top-left pixel of each block and discards the other three. libjpeg instead averages all four (see jcsample.c::h2v2_downsample).

On natural photographs the difference is negligible. On images with sharp colored edges — UI screenshots, rendered text, anti-aliased line art — point-sampling can stamp an entire 2×2 block with one outlier pixel's chroma, producing visible color fringing that box-averaging avoids.

On a mixed set of ~1,500 images (photographs, charts, UI screenshots, rendered text) encoded at quality 75, mean per-pixel-channel |Δ| against libjpeg-turbo's output drops from 0.26 with Nearest to 0.003 with Average.

What's in this change

• ChromaSubsamplingMethod { Nearest, Average } enum and Encoder::set_chroma_subsampling_method.
• A get_block_averaged path alongside the existing get_block, wired into both the interleaved and progressive encode paths. It's generic over stride, so 4:2:2, 4:1:1, etc. get the same treatment.
• The rounding bias alternates per output column (1,2,1,2,... for 4:2:0), matching libjpeg's dither so rounding error averages to zero across each row rather than accumulating.
• Four new tests: two unit tests on the averaging arithmetic and two encode→decode round-trips (interleaved and progressive).

Compatibility

Default remains Nearest. No change to output for any existing caller.

Related

Tangentially related to image-rs/image#2202 (JPEG encode quality), since image now uses this crate for JPEG encoding.