Fix JPEG scaled decoding parity for 0.11.2

[jcwal1516]

JPEG reduced-resolution decoding could differ from libjpeg-turbo because components used the same IDCT size, chroma was smoothed at unsupported scales or widths, and progressive/12-bit paths decimated full-resolution pixels. Apply per-component reduced IDCTs and the matching upsampling rules across the CPU and Metal paths for the planned 0.11.2 release.

Also fix 12-bit region slice panics, Metal region chroma context at MCU boundaries, CUDA CPU-backed surface dimensions, and the 12-bit scratch budget for live output rows. A 208-image libjpeg-turbo 3.1.4.1 golden matrix covers full images, regions, scale factors, restart markers, progressive images, supported 12-bit layouts, and tall narrow images. Existing unsupported 12-bit layouts still return NotImplemented; narrow subsampled images reject explicit device requests and use CPU decoding under automatic routing.

Validation:

• CPU golden matrix and decode-into integration suites: 88 passed.
• Reference geometry check: 1 passed. The 192 original JPEG inputs and reference samples were preserved when adding the 16 tall-image cases.
• WSI parity: 10 passed; real-GPU Metal matrix: 5 passed; CUDA-wrapper region tests on macOS: 6 passed.
• Tall 4×64 12-bit 4:2:2 and 4:2:0 fixtures reproduced MemoryCapExceeded before the budget fix and now match libjpeg-turbo, including region reads.
• Formatting and diff checks passed. Linux CUDA hardware validation and hosted PR checks remain pending; this PR does not publish a release.