tooling/libpng-oracle: read back text and colour chunks so metadata preservation is oracle-testable

[justin13888]

Summary

libpng_oracle::decode returns pixels plus IHDR fields and nothing else:

pub struct DecodedImage {
    pub width: u32, pub height: u32, pub bit_depth: u8,
    pub color_type: u8, pub interlace: bool,
    pub rowbytes: usize, pub pixels: Vec<u8>,
}

So a test can ask the reference reader "does this file decode, and to the same pixels?" but never
"what metadata did you read out of it?". For gamut-png's metadata-preservation path (#483, PR
#550) that is the whole claim: a re-encode must put every eXIf / iCCP / sRGB / cICP / gAMA / cHRM /
tEXt / zTXt / iTXt payload back, in the right chunk, with the right bytes. Today that is pinned by
gamut reading its own output — a round trip, which by construction cannot see a defect symmetric
across this crate's reader and writer — plus a third-party spot check done by hand.

Proposal

In tooling/libpng-oracle/src/lib.rs, extend DecodedImage with what libpng already parsed
during png_read_info / png_read_end:

• texts: Vec<OracleText> from png_get_text, each carrying keyword, text, and libpng's
  compression field — which distinguishes tEXt (PNG_TEXT_COMPRESSION_NONE), zTXt
  (_zTXt), and the two iTXt forms (PNG_ITXT_COMPRESSION_NONE / _zTXt) along with
  lang and lang_key. That is exactly the identity TextChunkKind records on the gamut side;
• iccp: Option<(String, Vec<u8>)> from png_get_iCCP;
• srgb: Option<u8> from png_get_sRGB, gamma/chromaticities from png_get_gAMA_fixed /
  png_get_cHRM_fixed, exif: Option<Vec<u8>> from png_get_eXIf_1;
• cicp, which libpng 1.6.43 does not parse — read it as an unknown chunk via
  png_set_keep_unknown_chunks + png_get_unknown_chunks, which is also how the caBX store
  could be checked.

Then crates/gamut-png/tests/preservation.rs can assert the reference reader sees what the
source carried, rather than gamut agreeing with itself.

Note libpng needs png_set_text / keyword handling to be Latin-1-correct for the text half of
this to be meaningful — that is #571.

Why not now

tooling/ was outside PR #550's manifest.

Related: #502 (bKGD/sBIT plus a warning count), #571 (Latin-1 text payloads).