Test the image replacement path

[psobot]

Why

file_utils' Data/ image-replacement branch had zero test coverage, and no test in the repo imported PIL at all. That's how Image.ANTIALIAS survived there for two years after Pillow 10.0 removed it (July 2023): every user's image replacement raised AttributeError, and nothing in CI noticed.

#59 fixed the call but added no test, so the same class of breakage would go unnoticed again.

What

tests/test_image_replacement.py drives the real path against tests/data/table.key:

• matching assets are actually rewritten
• the replacement is scaled down to each original size variant, so one source image can stand in for every size Keynote generated
• the original format is preserved
• a non-matching pattern leaves the Data/ assets untouched
• the output is still a readable Keynote archive, not merely a valid zip

No new binary fixtures — the replacement image is generated with PIL at test time.

These tests actually catch the bug

Rather than assume, I reintroduced Image.ANTIALIAS and re-ran them:

FAILED tests/test_image_replacement.py::test_matching_assets_are_rewritten
FAILED tests/test_image_replacement.py::test_replacement_is_scaled_to_each_original
FAILED tests/test_image_replacement.py::test_replacement_preserves_the_original_format
FAILED tests/test_image_replacement.py::test_output_is_still_a_readable_keynote_archive
4 failed, 1 passed

Coverage

keynote_parser/file_utils.py: 40% → 67%. Full suite 35 passed.

One line in that branch is still uncovered by design: the continue taken when a find string doesn't split into exactly two dot-separated parts. That's the #52 failure mode — a pattern like game(?:-\d+)?\.png is silently skipped. I've deliberately not pinned that behaviour down in a test, since it's a wart to fix rather than a contract to lock in.