Add Parallel research subagent to Pi

[georgeatparallel]

This adds a parallel-research agent to our existing @parallel-web/pi-extension package. Pi can hand off a web research question and get a finished answer with sources back as the child's normal output, including in JavaScript workflows.

Each research call makes one non-streaming POST /v1/responses request. There is no second synthesis pass, automatic retry, or background job. Pi keeps orchestration, waiting, branching, and artifact delivery. Cancellation aborts the local request on a best-effort basis; it does not confirm that research stopped on the server.

Default requests contain only the packaged research instructions and the latest task. Parent history, local context, session details, tools, and artifact paths stay with Pi. Trusted callers using low-level payload or header overrides own any changes to that boundary.

After publication, install pi-subagents and @parallel-web/pi-extension through Pi, run /login parallel, then use /run parallel-research <question>. Check auth status with /parallel-login. Use /logout and select Parallel to remove a stored credential; this does not unset PARALLEL_API_KEY.

Validation at 715c444:

• 147 local tests passed, including all 46 Pi extension tests. The 14 tests needing live API credentials were skipped.
• Lint, formatting, typecheck, build, package discovery, and a simulated version-only release bump passed. All six GitHub CI checks passed.
• The packed extension passed /run, JavaScript branching, and HTTP error, timeout, and abort checks using local HTTP fixtures. Requests excluded local runtime metadata, and Pi received cited output and saved the artifacts. The research probes used fixtures. Separate full-suite runs also exercised the existing live Search/Extract integration tests.

The package stays at 1.2.0 in this PR. The usual release PR will publish the update after merge. Public npm installation is still untested until publication.