gainmapmath: directly sample gain maps at scale 1

[gregbenz]

For a gain map at scale 1, the Shepard weights are [1, 0, 0, 0]: the result is just the first sample. The integer samplers still load four neighboring samples and combine them. This change reads the selected pixel directly for grayscale and RGB/RGBA maps, keeping the existing edge clamping, pixel strides, and / 255.0f conversion.

Performance

In a local 12 MP JPEG decode test with a full-resolution RGB gain map, this reduced elapsed time by 8.6-9.5% and process CPU time by 15.5-15.9%.

Run	Baseline elapsed	Patched elapsed	Baseline CPU	Patched CPU
1	132.7 ms	121.3 ms	307.8 ms	260.0 ms
2	135.3 ms	122.4 ms	312.6 ms	262.9 ms

Each value is the median of nine iterations after two warmups. The test used a deterministic synthetic 4000x3000 JPEG on macOS ARM64, Apple Clang 21, libjpeg-turbo 3.2.0, and a Release -O3 build with intrinsics enabled. The in-memory harness included decoder creation, setters, decode to linear RGBA half-float, and release; file I/O was outside timing. CPU time includes all process threads. These results measure HDR decoding, not a complete transcoding pipeline.

Tests

The added tests cover padded grayscale and RGB/RGBA rows, edge clamping, alpha-independent RGB sampling, and all 256 grayscale byte values. They also pass against the unmodified baseline.

• JPEG-only Release build with warnings as errors: 1,099 tests passed; 224 skipped by the suite.
• All 33 baseline/patched raw-output comparisons were byte-identical across linear, HLG, and PQ, including varied per-channel metadata and a gain map one row shorter than its base.
• ASan/UBSan: all 64 gain-map math tests and six focused CLI output comparisons passed with halt-on-error enabled. Leak detection was disabled because this macOS runtime does not support it; system libjpeg was not instrumented.