refactor(linux): finish remediation steps 31 to 39

[datlechin]

Why

The rest of Steps 31 to 39 of linux/docs/plan-2026-09-14-audit-remediation.md, after #3403. Tabs did not own their fetches, DDL came from one generic builder, closing could drop unsaved work, and pending grid edits followed column positions.

Change

• Table and structure tabs own sequenced, cancellable fetches over a shared schema catalog, and recover after a reconnect.
• DDL is planned per engine (renames, primary keys, identity, SQL Server defaults), keeping every column attribute; a DDL conformance suite runs on all five engines.
• Saves go through a ledger with Stop and unknown outcomes; close, disconnect and switching connection ask before losing edits or open editor transactions; tabs and drafts persist per connection.
• Grid edits follow edit order with undo, survive a column change by name, and edit long text from its full value; SQL completion reads the catalog.

Verified

fmt, both clippy runs and every unit and widget test pass; docker suites ran per engine for DDL and saves. Each data-integrity fix has a test that failed before it.

Not in this PR

• MySQL VerifyFull over TCP still checks the dialled host until sqlx gains a server-name override.
• A SQL Server stop during a quiet stretch after rows still drops the editor session (cancel_query fails and leaves the connection out of step when the attention ack comes in a separate message tiberius-rs/tiberius#459).
• PostgreSQL type changes reset SET STORAGE and SET COMPRESSION, which the catalog does not read yet.
• A bulk delete waiting on its confirmation keeps the row keys it captured before a primary-key change.