Horizontal predictor (one_byte_predict, all integer TIFFs with
Predictor=2): rewrite with explicit [u8; N] accumulator so LLVM unrolls
and keeps each channel in a register. For N=3 (RGB8), this replaces a
scalar 1-byte-per-iteration loop with a 2x-unrolled 3-register loop.
Float predictor (Predictor=3, f16/f32/f64): split deinterleave into
per-format functions with plane slicing up front; f32 batches 4 samples
through a u128 wide store to eliminate per-sample bounds checks. Thread
a reusable scratch buffer through expand_chunk to eliminate per-row
heap allocation.
Benchmarked on real 21MP photos and GeoTIFF DEMs (vs upstream main):
Strip RGB8 deflate+hpred: 253ms -> 224ms (-11%)
Strip RGB8 LZW: 405ms -> 370ms (-9%)
Tiled RGB8 deflate+hpred: 241ms -> 222ms (-8%)
Strip RGB f32 deflate+fpred: 726ms -> 527ms (-27%)
Gray f32 tiled 512: 290ms -> 166ms (-43%)
Copernicus DEM GeoTIFF: 239ms -> 152ms (-37%)
USGS DEM GeoTIFF: 284ms -> 198ms (-30%)
Lilith River
fa8a911f79