fix(sdk): release completed embedded request signals

[opencode-agent]

Problem

OpenCode Agent memory profiles show sustained growth in Request, Headers, AbortSignal, and array-buffer counts. A local red-capable repro using the actual OwnedFetch transport and Effect web handler retained 120/120 completed Requests after forced GC while the embedded host stayed open. The long-lived shutdown signal held the AbortSignal.any child; Effect’s abort listener on that child kept each handler Request reachable. The prior SDK signal-listener change in draft #51012 did not sever this parent edge.

Fix

Use a per-request controller with explicit source/shutdown abort forwarding. Remove both forwarding listeners when the response finishes or the handler rejects. This preserves caller cancellation, host-close cancellation, and streaming body cancellation while releasing completed requests.

Verification

• Regression test failed on origin/v2 with 120/120 Requests retained; after fix, no more than 3/120 survived GC in three consecutive runs (last async frames may remain live). Failure-path retention and caller/host/stream cancellation tests also pass (5 focused tests total).
• SDK typecheck, targeted Prettier, targeted oxlint, and diff whitespace checks pass.
• Existing transport.test.ts fails identically on both clean origin/v2 and this branch due an unrelated expected SessionNotFoundError vs actual NotFoundError mismatch.
• Full repo bun run check reaches typecheck, then is blocked by this Host’s Node v18 (@opencode/posts requires >=22.12); an earlier unconstrained run also hit the Host process/thread limit. Repo-wide lint passed. Neither limitation changes the scoped SDK checks.

This is a measured local retention fix, not a claim that production RSS has already fallen. No merge, deployment, or service restart was performed. Related: anomalyco/anomaly#260 and anomalyco/anomaly#623.

Requested by: @jlongster (James via Slack)

[kitlangton]

Superseded by #53613.