gamut-exif: to_bytes re-emits a thumbnail pointer whose bytes the report says were dropped

[justin13888]

What happens

ExifReader::read_thumbnail removes the JPEGInterchangeFormat (0x0201) tag from the stored 1st
IFD only when bytes were actually read:

let mut ifd = ifd;
if jpeg.is_some() {
    ifd.remove(ptr);
}

When the range was out of bounds — the case leniency exists for — jpeg is None, so the pointer
survives into the model and Exif::to_bytes writes it back out verbatim. The emitted blob then
claims a thumbnail at an offset that addresses nothing, while the ReadReport for the same parse
says the bytes were dropped.

Reproduction (executed)

A 1st IFD with JPEGInterchangeFormat = 0xFFFF and JPEGInterchangeFormatLength = 16, parsed
leniently and re-serialised:

report:            [Dropped { region: ThumbnailJpeg, tag: Some(513), offset: 65535, reason: OutOfBounds }]
pointer in model:  Some(65535)
re-parsed from to_bytes(): pointer still Some(65535), report still names the same drop

So the round trip preserves a dangling pointer rather than either dropping it or repairing it, and
a consumer of the emitted blob has no report to tell it the pointer is a lie.

Why it is filed rather than fixed

The behaviour is pre-existing: the same conditional removal is on master, byte for byte, and
predates the read report. Repairing it changes what the writer emits for a malformed input,
which is well outside the manifest of #419 (whose deliverable is the report). PR #522 documents the
interaction instead: the report claims completeness over the thumbnail bytes, and does name this
drop — it is the emitted blob, not the report, that is silent.

Options a fix should weigh

1. Remove the pointer unconditionally, so the model never carries an address it could not follow.
   Consistent with how the Exif/GPS/Interop pointer tags are already stripped, and with
   Thumbnail::from_jpeg, which deliberately does not store the offset because the writer
   synthesises it. Changes to_bytes output for a malformed input.
2. Keep the pointer but have the writer refuse to emit an offset it cannot satisfy.
3. Leave it and document it as a preservation guarantee — hard to defend, since the preserved value
   is knowably wrong.

Option 1 looks right, but it is a writer behaviour change and wants its own equivalence sweep.

Refs #419, #522.