feat(diagrams): add raster fixtures for pixel-only recovery

[MannXo]

Twelve raster fixtures. Ten redraw a graph that also ships as SVG, against the same ground truth,
so SVG, vector PDF and raster can be scored apart.

Raster is the container that keeps least. A vector PDF still holds path operators and positioned
glyphs; a PNG holds pixels, so the shapes and the text both have to be recovered before a graph
can be inferred at all.

Built by scripts/build_diagram_rasters.py, which mirrors build_diagram_pdfs.py and has the
same --check. Determinism comes free here. Neither engine writes a tIME chunk or anything
else derived from the clock, so two runs agree byte for byte with no normalisation step.

Two engines, named in each filename because the engine and not the source decides the pixels.
Mermaid cannot go through cairo for the reason you already documented for the PDF set. That
failure is silent, since the file is valid and only the text is missing, so
test_diagram_manifest.py now asserts that a fixture drawn from a <foreignObject> source is
built by Skia rather than leaving it to whoever edits the recipe next.

graphviz_cjk ships here. A PNG holds an image of text rather than the outlines that drew it, so
no font is redistributed and the fsType question does not arise, which recovers CJK, Hebrew and
Arabic coverage for the one container that had lost it. Flagging rather than assuming, since the
call is yours.

Everything is drawn at 2x. A caption set around 12px sits near the floor of what OCR reads
reliably at 1:1, and a benchmark should not be measuring a resampling artefact.

PNGs are bucket-managed and still need publishing. corpus.lock.json is unchanged, so the twelve
pin assertions fail exactly as the ten PDF ones already do on main. Files ready to hand over.

[MannXo]

the ci failures are not related to this PR, all failures stem from test_should_index_a_corpus_binary_that_the_lock_file_pins