[ikashapov]

Decode WMF and SVG images instead of detecting and skipping them

Closes #150. Based on main (b2c0174, v0.5.1); one commit, branch
issue-150-wmf-svg. Independent of the other open branches.

What renders now

Both formats used to be recognised as image parts and dropped — the space
reserved, nothing drawn, because Skia's built-in decoder reads neither.

WMF — render::wmf, built to render::emf's pattern

The issue named emf.rs as the model to follow, documented limits
included, and that is what this is: handle the single-embedded-bitmap
case, decline anything that needs full GDI record replay. Per [MS-WMF]:

• the optional 22-byte placeable header (§2.3.2.3) is sniffed by its
  0x9AC6CDD7 key and skipped — standalone .wmf files carry it,
  clipboard/RTF metafiles don't, so both spellings parse;
• META_HEADER (§2.3.2.2) is validated on its two fixed fields
  (HeaderSize = 9 words, version 0x0100/0x0300);
• the record walk scans the four DIB carriers: META_DIBSTRETCHBLT and
  META_DIBBITBLT — matched on their low byte, as §2.1.1.1 requires, and
  each guarded by the spec's own discriminator for their DIB-less second
  forms (RecordSize == (RecordFunction >> 8) + 3) — plus
  META_STRETCHDIB and META_SETDIBTODEV. SRCCOPY / DIB_RGB_COLORS
  only, like the EMF twin.

The payload is the same DeviceIndependentBitmap object EMF carries —
[MS-EMF] defines its bitmap records by reference to the [MS-WMF] objects —
so the DIB decoding moved out of emf.rs into a shared render::dib,
and grew what clip-art metafiles actually contain:

• palette-indexed 1/4/8-bpp depths (RGBQUAD tables, biClrUsed
  shortening, an out-of-table index painting black rather than sinking
  the bitmap);
• the 12-byte BitmapCoreHeader (§2.2.2.2) with its RGBTRIPLE palette;
• BI_JPEG / BI_PNG (§2.1.1.7), whose "pixel data" is one embedded
  stream handed to Skia's ordinary decoder;
• and the packed-DIB split (header + colour table + pixels contiguous,
  §2.2.2.9) that WMF records use, next to EMF's separate-slices shape.

Both metafile paths gain all of it. RLE and 16-bpp decline — drawing
nothing beats drawing garbage. The placeable magic also joins
ImageFormat::detect_by_magic, so an extension-less WMF part classifies.

SVG — svgBlip parsed, the fallback outranked, resvg rasterizing

Word embeds SVG as an asvg:svgBlip extension
({96DAC541-7B7A-43D3-8B79-37D633B846F1}, [MS-ODRAWXML]) beside a
rasterized PNG the main blip points at — kept, in the spec's own words,
"for backward compatibility" with consumers that cannot draw SVG. This
renderer now can, so:

• the extension parses into the model (Blip.svg_embed), tolerant of the
  other extensions Word writes into the same extLst;
• it survives the header/footer relationship remapping (a header's SVG
  id must follow its part's namespace);
• selection prefers it over the fallback — the reading LibreOffice 24.2
  chose too. Producers that point the main blip straight at an
  image/svg+xml part (pandoc without a rasterizer at hand) take the
  same decode path.

Why resvg and why a raster. Skia has its own SVG module, but
skia-safe's svg feature forfeits the prebuilt-binary cache on every
platform this project builds on — no published binary pairs svg with
embed-freetype, and none carries it without a GPU backend or webp in
tow — turning every cold build into a from-source Skia build. resvg
with default features off is pure Rust (~40 crates, seconds to build,
MIT/Apache). The raster is produced at the painter's existing
per-placement downsample target, so an SVG comes out crisp at exactly
--image-dpi and the size-keyed bitmap cache keeps differently-sized
placements apart; alpha stays premultiplied end to end. The one recorded
limit: <text> inside an SVG is not rendered — the text feature would
pull a second font stack (fontdb/harfrust) beside Skia's, and
Office-inserted SVGs are overwhelmingly icons and diagrams.

Tests and fixtures

• test-files/wmf-image.docx + test-files/svg-image.docx (new, by
  scripts/make_issue150_fixtures.py, written red-first). The WMF wraps
  a hand-built placeable metafile around a 2×2 four-colour DIB. The SVG
  fixture makes the selection observable: the fallback PNG is blue, the
  svgBlip's SVG is red — which part was picked is the drawn colour — and
  a second picture carries a no-fallback direct-SVG blip.
• tests/wmf_svg_images.rs pins the pipeline: the extension in the
  model, both parts' formats in the media table, both draw commands
  resolving to the SVG parts, and a converted PDF that embeds an
  /Image XObject where it used to embed none.
• Unit tests: the WMF record offsets (both blt variants over one DIB —
  equality pins both transcriptions), the DIB-less-form discriminator,
  placeable/bare parity, non-SRCCOPY declines; the shared DIB decoder's
  palette packing (nibble and bit order pinned on asymmetric patterns),
  core header, BI_PNG passthrough, undersized-table behaviour, RLE
  decline; resvg rasterization at the display target, intrinsic size,
  transparency, malformed-input decline; and the EMF suite re-based onto
  the shared decoder unchanged in behaviour.
• README: the WMF row flips ❌ → ⚠️ (same single-bitmap contract as EMF),
  the SVG row ❌ → ✅, and EMF's row records its widened DIB support.

cargo fmt --check, clippy --all-targets -- -D warnings,
RUSTDOCFLAGS="-D warnings" cargo doc --no-deps and the full 45-target
test suite pass (macOS, pinned 1.95.0). New dependency: resvg 0.48,
default features off, raster-images + svgz only — no C code, no
build-script downloads, MSRV 1.85 against the repo's pinned 1.95.