fix(update): snapshot busy WAL databases instead of refusing admission (#201327)

Artifact-preserving SQLite inspection copies the database bytes and
compares them until they hold still. A live writer that checkpoints during
every copy (a busy Gateway agent turn on a large agent database) exhausts
all ten attempts, so update admission refused with "SQLite source did not
stabilize (reason=wal-copy-unstable, attempts=10)".

When every attempt loses to committed writes on a complete WAL family
(main, WAL and index present, no rollback journal), take one SQLite read
transaction through the existing online-backup owner instead. Only a live
connection can keep checkpointing, so its WAL index is current: the read
cannot recover it or change the 96-byte commit header that update captures
compare, only reader slots after it. Idle, crashed and index-less families
never churn and keep the byte-neutral copy; rollback-journal families keep
refusing because a native reader would block their writer.

Refs #200791