fix: use inverse CDF for histogram matching

[teddytennant]

Problem

match_histogram_mut produced suboptimal mappings for simple cases such as the one in #783:

let mut input_image = gray_image!(type: u8, 0, 1, 2, 3, 4, 5, 6, 7);
let target_image = gray_image!(type: u8, 30, 130, 30, 130);
match_histogram_mut(&mut input_image, &target_image);
// was: 29, 30, 30, 30, 129, 130, 130, 130
// want: 30, 30, 30, 30, 130, 130, 130, 130

The old result invents gray levels (29, 129) that never appear in the target, so the matched histogram is farther from the target than necessary.

Root cause

histogram_lut chose, for each source level, the target level whose cumulative fraction was nearest in absolute distance. When the target CDF jumps (for example from 0 at level 29 to 0.5 at level 30), source fractions just above 0 are closer to the pre-jump empty bin than to the level where mass actually appears.

Fix

Use the standard discrete inverse CDF:

lut[i] = min { y | target_cdf[y] >= source_cdf[i] }

Comparisons are done with integer cross-multiplication to avoid floating point. Empty source or target images leave the LUT as zeros.

Tests

• Updated test_histogram_lut_gradient_to_step_contrast for inverse-CDF expectations (maps onto 30/130 rather than 29/129).
• Added test_match_histogram_mut_maps_to_target_levels from the issue example.
• Regenerated tests/data/truth/elephant_matched.png for the regression suite.

cargo +nightly test --lib histogram
cargo +nightly test --test regression test_match_histograms

Fixes #783