Progressive encoding Q10 size diff ~2.8% vs C++ jpegli

[lilith]

Bug

Progressive encoding at Q10 produces ~2.8% larger files than C++ jpegli reference, exceeding the 2% tolerance threshold. Affects all subsampling modes.

Tests:

cargo test --release -p zenjpeg --features __ffi-tests --test quality_matrix -- test_ycbcr_420_progressive --ignored
cargo test --release -p zenjpeg --features __ffi-tests --test quality_matrix -- test_ycbcr_422_progressive --ignored
cargo test --release -p zenjpeg --features __ffi-tests --test quality_matrix -- test_ycbcr_440_progressive --ignored
cargo test --release -p zenjpeg --features __ffi-tests --test quality_matrix -- test_ycbcr_444_progressive --ignored

Symptoms:

• YCbCr 4:2:0 Progressive Q10: Size diff 2.8% (rust=117109, ref=113877)
• YCbCr 4:2:2 Progressive Q10: Size diff 2.9% (rust=128954, ref=125348)
• YCbCr 4:4:0 Progressive Q10: ~2.8%
• YCbCr 4:4:4 Progressive Q10: ~2.8%
• Higher quality levels (Q50+) are within tolerance

Context: Pre-existing issue surfaced by enabling __ffi-tests in CI. The size difference is at very low quality (Q10) only, likely from Huffman table or DCT rounding differences.

File: zenjpeg/tests/quality_matrix.rs:691