Decode palette PNGs instead of dropping them

[mantissaman]

Closes #17.

The bug

decode_image mapped PNG colour types to a channel count and sent anything unrecognised to a catch-all:

let channels: usize = match color_type {
    0 => 1, 2 => 3, 4 => 2, 6 => 4,
    _ => return None,
};

Colour type 3, indexed colour, landed there. Every caller treats None as nothing to draw, so an indexed PNG was silently omitted from both PDF and PNG output.

These are not unusual. Most tools default to a palette when an image has few colours, and Word embeds whatever it is handed. The file attached to #3 carries a 640x480 8-bit colormap PNG, which is why that document looked like it had no image even after #10 corrected the anchor position.

The fix

PLTE supplies the palette and tRNS supplies per-entry alpha when present. Each index is expanded to RGB or RGBA after unfiltering, which is where the existing code already has a clean one-byte-per-pixel scanline to work from.

Three decisions worth calling out:

• An index past the end of the palette is malformed input. It draws black rather than discarding the whole image, so one bad byte does not cost you the picture.
• A missing PLTE is still rejected, because there is nothing to expand against.
• Fully opaque images carry no alpha channel, matching what the RGBA path already does, so PDF viewers do not get an SMask they do not need.

Bit depths below 8 are still rejected by the existing bit_depth != 8 check. Sub-byte indexed images stay unsupported rather than silently wrong, and that is now a deliberate limit rather than an accident of the catch-all.

Verified

Four tests build indexed PNGs from scratch and cover palette expansion, tRNS alpha including entries it does not mention, the missing palette case, and the out-of-range index case.

Rendering the file from #3 with this and #18 applied together, the image finally appears, in the right place, with its transparency intact. The page went from 19 KB to 70 KB of PNG output.

Checks

cargo fmt --all --check, cargo clippy --workspace --all-targets -- -D warnings and the full suite all pass. 347 tests, up from 343, with four added here.

python3 scripts/hash_harness.py --check also passes, 28 entries matching. Worth stating plainly why: none of the harness fixtures contains an indexed PNG, so the baselines are genuinely unaffected. This change does alter output for any document that does contain one, which is the entire point of it.