fix: close Hyperdrive client after query failures

[kaijakagi-sys]

Problem

A failed Hyperdrive query skips sql.end() in the base implementation. Moving cleanup into finally fixes that leak, but an awaited cleanup rejection can then hide the original query rejection.

Change

• Always close the Hyperdrive client after query success or failure.
• Preserve the original query rejection if cleanup also fails, including an undefined rejection reason.
• Keep cleanup-only errors visible and preserve background cleanup through executionContext.waitUntil().
• Add ten focused lifecycle regression tests with a mocked Postgres client.

Verification

Base: bb227352135051a0dd50190d67c3e82b03f8e485. Updated head: 958168e4086f5b1b0f0dbe75400193eb06c5c93a.

• Original leak: two rejection-cleanup tests failed before the initial fix.
• Dual-failure regression against previous draft head bec45944ec3464ec1e2364a143c213f659bf4ed7: 2 failed, 8 passed.
• Updated vitest run src/operation.hyperdrive.test.ts src/operation.test.ts: 21 passed.
• Full suite: 161 passed, 4 RLS failures. Previous draft head: 155 passed, the same 4 failures.
• Typecheck: the same 11 diagnostics at both revisions, apart from shifted source locations.
• Prettier and git diff --check pass.

These are local Node 22.23.2 / Vitest 2.1.8 tests; no live Hyperdrive deployment was tested. The repository CI uses Node 20, which was not tested locally.

Short demonstration

Recorded local regression results, not a deployment recording:

hyperdrive-regression.mp4

Scope

Related to #71. This is a narrow regression contribution, not completion of the repository-wide 75% coverage goal. #153 covers successful Hyperdrive cleanup; this change covers query-rejection cleanup and simultaneous failures. No reward has been agreed for this slice; keeping this draft pending scope/eligibility confirmation.