fix: Read files in raw mode so flobs preserve every byte

[joethorley]

Closes #24

Problem

flob() read files with readBin(what = "integer", n = size), which:

• dropped the trailing 1-3 bytes of any file whose size is not a multiple of 4, corrupting formats such as .tar.gz and .xlsx
• returned NA_integer_ for any 4-byte-aligned chunk equal to 00 00 00 80, so vld_flob() returned FALSE and unflob() errored depending on file content

Changes

• flob() reads files with readBin(what = "raw").
• vld_exint() / chk_exint() accept raw or integer elements, so legacy integer flobs (e.g. those already stored in databases) remain valid and unflob.
• The not-NA check is dropped, since NA is a legitimate bit pattern in legacy integer flobs. This also removes the call to the deprecated chk::chk_no_missing().
• flob_obj is regenerated as a raw flob. The internal flob_old, flob_noname and slob_obj objects stay integer-based to keep legacy coverage.
• The DESCRIPTION, README and ?flob descriptions now say raw-mode instead of integer-mode.
• Tests added for byte-for-byte round trips of 0-9 byte files, the NA bit pattern, legacy integer flobs containing NA, and the new chk_flob() error message.

Notes

• Bytes already lost from existing integer flobs cannot be recovered.
• dbflobr has its own copy of vld_exint() / vld_slob() (R/vld.R) that requires integer elements. It will need the same change to accept raw flobs through its slob path.
• The remaining test failures in test-chk.R, test-ext.R, test-flobr.R, test-path.R and test-tools.R also occur on main. They come from newer chk error messages and tools::file_ext() behaviour (CI: file-extension test failures on main (pre-existing) #19).
• The issue's tarball example now round-trips with an identical md5sum.

🤖 Generated with Claude Code