Fix nativeRaster to be row-major so bitmaps serialize correctly

[billdenney]

pdf_render_page(), pdf_image_bitmap(), pdf_image_rendered() and pdf_text_obj_rendered_bitmap() return a pdfium_bitmap that inherits from nativeRaster. The C++ writers filled the backing integer buffer column-major (the R matrix default), but a nativeRaster's buffer must be row-major: element k is the pixel at (row = k / width, col = k % width).

Any consumer that trusts the nativeRaster class and reads the buffer directly -- png::writePNG(), grid::grid.raster(), R's graphics engine -- therefore read every row from column-major storage and sheared it sideways ("stride-streak" garble). Downstream packages worked around this by re-decoding via as.array() or by reshaping inflated image streams by hand. The pixels PDFium produced were correct all along; only the R-side buffer layout was non-conformant.

Fix: emit the buffer row-major so the object is a genuine, conformant nativeRaster. png::writePNG(bmp) and grid::grid.raster(bmp) now render correctly with no reshape, matching png::readPNG() / magick::image_read() in dims, RGBA channel order and 0..1 range.

• Add src/native_raster.h with shared row-major fill helpers (fill_bgra_rowmajor for the always-BGRA render path; fill_bitmap_rowmajor for the Gray/BGR/BGRx/BGRA image-object path), honoring FPDFBitmap_GetStride() per row.
• render.cpp, images.cpp, tier3_extras.cpp: use the helpers (removes three near-duplicate copies of the format switch; the tier3 writer's prior "row-major" comment was wrong -- it used m(y, x), which is column-major).
• as.array.pdfium_bitmap() / as.raster.pdfium_bitmap(): unpack the row-major buffer with matching byrow reshaping.
• Tests: add nativeRaster-conformance regression tests for pdf_render_page() and pdf_image_bitmap() asserting writePNG(bmp) == writePNG(as.array(bmp)) byte-for-byte, plus a quadrant round-trip; update the hand-built column-major plot fixture to build a conformant row-major nativeRaster.

Verified: page 13 of a corpus scan (previously OCR'd to nothing through png::writePNG(bmp)) now OCRs its value axis correctly and is pixel-equal to the as.array() path and to magick/gs; full test suite passes (0 failures, 3021 assertions); a normal text page and a vector page show no regression.