fix(table): closing a table tab releases its lease without interrupting the shared session (#1215)

[ZhuchkaTriplesix]

Refs #1215

What was wrong

Closing a table browser tab called cancel(interruptIfBusy: true). Despite the name, the PostgreSQL, MySQL and SQLite table delegates interrupted the readOnly and tableWrite sessions of the database unconditionally. Those sessions are the same objects the tree, the other table tabs, the Relations view and MCP use, so closing one tab dropped the connection under the others.

Changes

• cancel in the three table delegates no longer closes a pooled session. A statement still running finishes or times out on its own (the timeouts from fix(db): a timeout or a server-side cancel force-closes a shared session, and a queued statement can cancel another caller's #1216 apply).
• GenericTableView.dispose no longer calls cancel(interruptIfBusy: true); the delegate's dispose releases the lease.
• A widget test: closing a table tab calls dispose once and cancel never.

Behaviour change

• A table query still running when its tab closes is no longer cut on the server; it runs until it finishes or times out. Its result is dropped.
• The ER diagram's table adapter forwards cancelQuery to cancel, so it no longer interrupts anything. The Relations view does not call it today.

Not in this PR

• The sequence and routine views interrupt the read-only session on dispose too (postgres_sequence_view.dart:56, postgres_routine_view.dart:64). Same bug, not part of fix(db): closing a table tab force-closes the read-only session other views are using #1215's list; a follow-up.
• A test that closing one of two table tabs does not break the other needs a fake pool connection under the table delegate; the widget test covers the call, not the shared session.

Not verified locally

Tests were not run locally, per the project rule.

[github-actions]

Grid scroll benchmark

metric	base	PR	change
p50	8.83 ms	9.67 ms	+9.5% ⚠️
p90	14.32 ms	16.55 ms	+15.6% ⚠️
p99	32.34 ms	38.14 ms	+17.9% ⚠️
stutters	346.00	335.00	-3.2%

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