Use native sequential archive input for TAR and CPIO

[Hawkynt]

What changed

• Override the universal forward-only archive input surface for TAR and CPIO using their existing sequential readers.
• Keep the overrides explicit on IArchiveFormatOperations, so concrete descriptor public API/documentation does not grow just to specialize a default interface member.
• Rewind seekable callers for semantic parity with the default adapter, while non-seekable callers are consumed directly.
• Add regression tests that append trailing junk and prove the source is not fully read/spooled before parsing.
• Add a TAR single-entry test proving OpenEntryStreaming can return the first entry without consuming the remainder of the archive.

Why these formats

POSIX ustar/pax is intrinsically sequential: a 512-byte header is followed by that entry's data blocks, with two zero blocks terminating the archive. CPIO is likewise a sequence of header, pathname, payload records ending at TRAILER!!!. Their existing readers already support non-seekable sources, so the generic temporary-file compatibility spool added by #373 is unnecessary overhead for these formats.

AR was reviewed but deliberately not changed: the current reader materializes ordinary members and depends on seeking for oversized members, so a native streaming override would not yet be uniformly safe.

Reference material

• POSIX pax/ustar interchange description: header followed by file data records and two 512-byte zero terminators.
• POSIX/FreeBSD cpio format descriptions: sequential header/name/data entries terminated by TRAILER!!!.
• libarchive read API architecture: streamable TAR/CPIO-style readers can operate without a seek callback; seekability remains a separate capability.

No external source code was copied or translated. No new dependencies were introduced.

Validation

• Branch is based on current main at 7bf197676b1a33039caaddb9f79da27ffdb7328c.
• Full CI pending.