fix(tiff): hardening: prevent OOB read, overflow protection

[lgritz]

In the raw-strip read path of read_native_scanlines(), a failed TIFFReadRawStrip() returns csize < 0. Bail out of the strip loop for this case before running the decompress, matching the existing tiled path (read_native_tiles already breaks on this condition), and guard the leftover-scanline loop with ok so it doesn't run after the bail.

Add regression seed to the TIFF corpora -- a valid DEFLATE TIFF with StripOffsets[0] pointing past EOF that the nightly ASan fuzz job exercises via full-image reads. The reader now rejects it with a clean error instead of reading out of bounds.

Also size scanline scratch with 64-bit arithmetic, for overflow protection.

Assisted-by: Claude Code / Claude Opus 4.8

[lgritz]

This is the first PR in a new initiative I'm taking to proactively look file format by file format and try to uncover potential bugs before the fuzzers or the bad guys even find them. So expect another series of PRs. I will try to organize them into one per format.

[lgritz]

Near as I can tell, the "hobbled" test failures are unrelated to this PR and are happening (nondeterministically) on several branches/PRs. I think it's some kind of GH configuration thing, am looking at it separately.