Feature/content aware quality map

[bfalls]

• Added a saliency API surface for per-8x8 block importance scores with configurable edge/contrast weighting and optional debug outputs.
• Implemented CPU saliency calculation (luma extraction, blur-based contrast, Sobel edges, block aggregation, percentile normalization) plus heatmap/CSV debug writers.
• Registered the saliency header and source in the Visual Studio project and filters.
• Applied per-block coefficient scaling in the CPU DCT/quantization loop so quality maps bias bits toward important regions without altering global quant tables.
• Derived saliency-driven block scaling from quality-map settings, including strength/min/max mapping, debug-aware saliency computation, and milder chroma scaling before invoking CPU compression passes.
• Added an optional per-block inverse scale parameter to the GPU channel compression API to support quality map inputs without altering existing defaults.
• CUDA DCT/quantization now accepts a device block-scale map and applies per-block coefficient scaling before quantization for standards-compliant quality-map control.
• The GPU compression path computes saliency-driven block scales on the host, uploads per-channel maps (with milder chroma influence), and threads them through the CUDA launches.
• Added helpers to auto place quality-map debug artifacts during compare runs, capture output sizes/timings/PSNR, and print a condensed compare table while avoiding duplicate GPU debug writes.
• Documented a quality-map demo command that emits heatmaps and comparison output details in the README
• Ensure the quality-map compare summary table always prints during compare runs, even when the GPU path is unavailable, filling in placeholder values when needed.
• Enabled self-hosted builds, cached vcpkg and use local CUDA 13.0.