Support zip_fseek() for stored entries encrypted with WinZip AES

[fdegros]

Summary

WinZip AES uses AES-CTR, which has no chaining dependency between blocks - decrypting block n only requires the counter value for that block, not any preceding ciphertext or plaintext. This adds _zip_winzip_aes_seek() to jump the cipher's internal counter/keystream state directly to an arbitrary byte offset, and wires ZIP_SOURCE_SEEK/ZIP_SOURCE_TELL into the WinZip AES decode source accordingly, so zip_fseek()/zip_file_is_seekable() now work for STOREd (uncompressed) entries encrypted with WinZip AES.

The one real constraint is the mandatory HMAC-SHA1 authentication code, which is computed over the ciphertext sequentially from the start and can't be verified for a chunk read out of order. This is handled exactly like zip_source_crc.c already handles the same situation for the CRC of a stored, unencrypted entry: the amount of ciphertext fed to the HMAC contiguously from position 0 is tracked, and verification is silently skipped (not failed) if a seek ever leaves a gap that's never filled in. A full, contiguous read from the start still verifies the HMAC exactly as before, including still catching a tampered file - only reads that actually skip around lose that guarantee, the same tradeoff already made for CRC.

Also updates zip_fseek()'s man page: it previously said compressed or encrypted data always errors out, which is no longer accurate.

Fixes #336.

Test plan

• ctest (196/196 tests pass)
• New regression tests: winzip_aes_seek.test, winzip_aes_seek_bad_hmac_full_read.test, winzip_aes_seek_bad_hmac_partial_read.test
• Manual verification of correct decryption at arbitrary (including non-block-aligned) offsets, cross-checked against full sequential reads, both on synthetic fixtures and on a real ~1GB AES-256 entry created independently with 7z
• Manual verification that a full sequential read of a file with a tampered authentication code still fails, while a seeked/partial read of the same tampered file succeeds silently (matching the documented, intentional tradeoff)
• clang-format --dry-run --Werror clean on all touched C/H files
• Man pages regenerated (.mdoc/.man/.html) for the touched page

🤖 Generated with Claude Code