perf(ai): clean up stream abort listeners

[JustYannicc]

What changed

• Route Raycast AI.ask() stream completion, stream error, electron.aiAsk rejection, and caller abort through one cleanup helper.
• Track the caller AbortSignal listener per active stream so terminal paths detach the listener before deleting _activeStreams.
• Add a focused node:test harness for retained abort-listener counts across success, stream error, aiAsk rejection, and explicit abort.

Why

Extensions can reuse a long-lived AbortSignal for many successful AI asks. The stream map was cleared on terminal paths, but the inline abort listener remained attached to the caller signal, retaining the request closure until that signal was eventually aborted.

I checked a bounded timeout/finalizer for never-finished renderer requests. This PR keeps behavior intact by not adding a new request timeout; it centralizes cleanup for every existing terminal path and leaves long-running streams unaffected.

Compatibility impact

• No API behavior removed.
• AbortSignal cancellation still calls electron.aiCancel and rejects with Request aborted.
• Successful and failed streams resolve/reject as before, with listener cleanup added.

How tested

• Before fix, the new harness reported retained listener counts on all terminal paths: active=1, added=1, removed=0.
• After fix: active=0, added=1, removed=1 for success, stream error, aiAsk rejection, and explicit abort.
• node --test scripts/test-raycast-ai-abort-listeners.mjs
• node --test scripts/test-use-ai-stream-batching.mjs scripts/test-raycast-ai-abort-listeners.mjs on the integration stack before the clean cherry-pick.
• ./node_modules/.bin/vite build
• ./node_modules/.bin/tsc -p tsconfig.renderer.json --noEmit --pretty false still fails on the clean origin/main baseline with unrelated existing renderer errors.

Stack validation

• Clean PR branch from origin/main; only this fix commit is included.
• Does not include perf(runtime): reduce extension and menubar churn #530, perf(ai): batch streaming and status updates #531, fix(validation): restore project checks #532, perf(renderer): cache icons and virtualize launcher surfaces #533, perf(mac): reduce native polling and async IO stalls #534, or perf(search): reduce indexing and ranking overhead #535. Those remain separate upstream performance/typecheck stack PRs for consolidation.

[JustYannicc]

Superseded by #556, which consolidates this AI/native media performance batch.