fix: prevent false commit success after DM8 restart

[skhe]

Problem

A real DM8 restart while a DECIMAL(30,8) insert was uncommitted made commit() return success even though a new connection found zero rows. The bridge executed statements and COMMIT through sql.DB, allowing them to use different physical connections after a failure.

Change

• Pin manual transactions, including direct/prepared statements and COMMIT/ROLLBACK, to one physical connection.
• Keep autocommit connections eligible for visible-error recovery. Runtime autocommit changes update the DM wire-protocol flag and invalidate statements prepared under the previous mode.
• Add an isolated ARM DM8 restart regression to the GitHub CI gate. A lost manual transaction must report an error; an already-open autocommit connection must recover.
• Update the timeout contract: a manual transaction connection whose socket was closed must be discarded and reopened.

Verification

• Before fix: commit_after_restart=success; persisted_rows=0 on local ARM DM8.
• After fix: restart verifier passed; lost commit rejected and autocommit connection recovered.
• Dedicated local DMPYTEST account: 193 real-database cases passed.
• Targeted autocommit toggle and timeout regressions passed; both Go modules tested; workflow YAML and patch consistency checks passed.