优化qrcode: improve QR code parsing accuracy and speed

[Copilot]

QR code parsing had two core problems: a scoring function that gave wrong signal to the perspective matrix optimizer, and an exhaustive matrix search that made ~12,750 expensive evaluations per decode regardless of whether they helped.

Accuracy

Fix scoreArea scoring logic — the previous implementation rewarded black pixels in white areas and never scored black pixels in black areas, making the score a poor proxy for matrix correctness:

// Before: only white areas ever scored; black areas always returned 0
if qr.greyPixels[index] != 0 {
    if !isBlack { score++ }  // black pixel in white area — wrong
} else {
    if !isBlack { score++ }  // white pixel in white area — correct
}

// After: reward any pixel that matches its expected color
pixelIsBlack := qr.greyPixels[index] != 0
if pixelIsBlack == isBlack {
    score++
}

Speed

Replace autoAdjustmentMatrix exhaustive search with coordinate descent — the old algorithm tried all 255 index-combination × 50 scale steps against a frozen baseline matrix (improvements never compounded). Replaced with multi-scale coordinate descent (scales: 0.5 → 0.005) that updates the base on every improvement, converging in ~100 evaluations vs 12,750. The 256-row autoAdjustmentMatrixChangeIndex static table is removed.

Convert fillRegion recursive flood fill to iterative — eliminates the hard depth cap of 1024 that silently dropped pixels in large regions; only pushes the start of each new run onto the stack rather than every pixel.

Pre-allocate rowAverage in binarization — single allocation reused across all rows instead of one allocation per row.

Benchmark (BenchmarkTestNX)

	Before	After
Time/op	99.6 ms	69.2 ms (−30%)
Memory/op	118 MB	100 MB (−15%)
Allocs/op	19,591	5,252 (−73%)

Safety

Added len(result) >= 3 guard before UTF-8 BOM detection in read8BitByteData to prevent a potential index-out-of-range panic on short payloads.