fix(db): statements run one at a time per session; a server cancel keeps the session (#1216)

[ZhuchkaTriplesix]

Refs #1216

What was wrong

• Every TimeoutException closed the pooled session, including a server-side cancel (SQLSTATE 57014). The driver reports that cancel as a PgException, and the session stays healthy; closing it dropped the tree, the table tabs, the Relations view and MCP together.
• The postgres driver starts a statement's timeout before the statement gets its turn behind another one on the same session. A statement that waited behind a slow one could cancel the slow one through the server.
• The MySQL client waits for the previous statement by polling and gives up after 10 s, which closed the session for every caller.

Changes

• PostgreSQL: a PgException from a statement keeps the session; a bare TimeoutException closes it, since the protocol may be out of step.
• PostgreSQL and MySQL: statements on one session run one at a time in our own queue. The driver's timer therefore starts when a statement starts, and no statement is cancelled for another one.
• executeWithTimeout passes its timeout to the statement instead of wrapping the whole call, so time in the queue does not count.
• Tests: a server cancel keeps the session; a bare timeout closes it; two statements on one session run in order (PostgreSQL and MySQL); a timed-out MySQL statement closes the session. The existing postgres timeout test now drives the real path through a test seam.

Not in this PR

• A timed-out MySQL statement still closes its session. The client cannot cancel a running statement; KILL QUERY on a second connection would keep the session, and is a follow-up.
• A waiting statement has no separate "waited too long for the connection" error; it waits in the queue, bounded by the statement's own timeout once it starts.
• A result read as a stream (iterable: true, MySQL SQL editor) is released from the queue before its rows are read; the client's own wait still applies to the next statement, as before.
• SQLite is not changed.
• The sequence and routine views (fix(db): closing a table tab force-closes the read-only session other views are using #1215 follow-up) are not changed.

Not verified locally

Tests were not run locally, per the project rule.

[github-actions]

Grid scroll benchmark

metric	base	PR	change
p50	14.34 ms	9.21 ms	-35.8%
p90	19.83 ms	22.66 ms	+14.3% ⚠️
p99	45.76 ms	46.27 ms	+1.1%
stutters	318.00	309.00	-2.8%

Informational only (threshold 5%). Shared CI runners are noisy; re-run before trusting a single result.