fix(scanner): read the luminance plane at its own row stride

[bmc08gt]

CodeExtractor built the scanner's input from CVPixelBufferGetBytesPerRow, which reports the buffer-level stride rather than plane 0's. On a planar 420YpCbCr8 buffer that is roughly 1.5x the luma row -- 2884 for a 1920-wide frame -- so the pointer handed to the scanner did not describe the bytes behind it.

kikCodeScan does memcpy(greyscale.data, image, height * width): it reads exactly width * height bytes from the front of whatever it is given and assumes they are a tightly packed greyscale image. CoreVideo aligns plane 0 to 64 bytes, so any width that is not a multiple of 64 arrives padded (1440 -> 1472, 1000 -> 1024), and those frames were being fed skewed rows.

This reads the plane-level geometry instead -- CVPixelBufferGetWidthOfPlane / GetHeightOfPlane / GetBytesPerRowOfPlane(buffer, 0) -- and strips row padding when the plane is padded. When rowStride == width the plane is already what the scanner wants, so it passes through with bytesNoCopy and no per-frame allocation.

Two related fixes fall out of the same path:

• The zero-copy Data escaped its CVPixelBufferLockBaseAddress scope. The base address is only guaranteed while the buffer is locked, so this was a latent use-after-unlock. The scan now runs inside the lock via withLuminanceSample.
• KikCodes.scan(data:width:height:) dropped the result of the quality overload it delegates to -- a missing return in Code.mm -- so the three-argument entry point always returned nil.

Cross-platform

Android had the same class of bug in the other direction: it guarded its unpadding with pixelStride != -1, which is vacuously true for a YUV_420_888 Y plane, so it ran a per-pixel copy on every frame. Two independent implementations of "is this plane packed?" produced two independent bugs.

The rule now lives once, in :libs:codes:kikcode as LuminancePlane, which SharedCoreKit already exports: code-payments/code-android-app#1303.

That version is not published yet, so this PR still carries a Swift-side implementation. CodeScanSweepTests.packingRuleMatchesSharedRule pins it to the Kotlin one byte-for-byte over the same geometries and the same fill until the pin moves to SharedCore 0.3.2, at which point the Swift copy is deleted and the test becomes a straight regression test.

Tests

CodeScanSweepTests is the iOS half of the harness mirroring Android's KikCodeScanTest. Both platforms feed the same C++ scanner -- the eight sources under CodeScanner/src are byte-identical to Android's vendor/kik/scanner/src/main/cpp -- so the pixel-buffer glue is the only place the two can disagree, and that glue is what these cover. It renders real codes, pushes them through real CVPixelBuffers at packed and padded widths, and checks that packed planes are genuinely not copied.