Reduce managed coupon barcode decoding CPU and latency

[Szer]

The managed barcode fallback repeatedly copies and rotates overlapping image regions. Use shared luminance views and SIMD row binarization to reduce that work, then decode independent crops with separate readers and a concurrency limit derived from Environment.ProcessorCount. The full-image fast path and original first-success crop priority are preserved; CPUs without SIMD use the scalar path.

In an isolated ARM64 comparison on one difficult image, five warmed paired runs reduced median managed-fallback time from 3.17 s to 1.27 s (60%) and process CPU time from 2.72 s to 1.58 s (42%). This is a targeted benchmark, not a production-wide latency estimate.

Validation:

• 264 hermetic HTTP integration tests passed.
• 40 OCR tests passed normally and again with hardware intrinsics disabled and one reported processor.
• Final implementation matched the unchanged managed decoder on all 2,201 offline corpus images, plus 71 selected cases on ARM64.
• Added coverage for low-contrast fallback recognition, concurrent calls, and patterned images with awkward widths; mandatory comment gate passed.

Recognition outputs are preserved, including existing limitations. The adapted ZXing threshold algorithm includes its attribution and Apache license.