feat: trace Portkey chat completions

[KarthikS (karthiksenv)]

Portkey's chat.completions.create(...) calls currently produce no Braintrust generation spans. This adds the normal Orchestrion configuration, a Portkey channel/plugin, and wrapPortkey(client) for portkey-ai v3 chat completions.

Both non-streaming and stream: true calls record messages, choices, model metadata, token usage, errors, and Portkey provenance. Streams aggregate text and tool-call arguments and record time to first token. Gateway configuration and HTTP options remain separate from logged metadata. Combining the wrapper with the automatic hook emits one span, and portkey / portkey-ai disable aliases are supported.

Refs #2482. This change covers the primary chat-generation API; prompt management, legacy text completions, and other Portkey APIs are outside this PR's scope.

Validation on Windows with Node 24.18.0, pnpm 10.33.0, and the official portkey-ai@3.1.0 package:

• Replaced the registration with the original main version: all four original automatic-tracing regression cases failed because no Portkey spans were emitted. Restored the fix afterward.
• 156 focused tests pass, including 21 Portkey cases across manual, automatic, and combined instrumentation. These use the real installed SDK with stubbed HTTP and exercise non-streaming output, SSE aggregation, tool calls, request-option forwarding, parentage, errors, early stream exit, stream failure, disable aliases, and wrapper idempotence.
• Source and test TypeScript checks pass; the SDK build (including declaration generation) passes.
• Packaged Node --import .../hook.mjs smoke test emits one Portkey span; disabling portkey-ai emits none.
• Repository formatter and git diff --check pass. ESLint reports no errors (two type-assertion warnings in the wrapper; test/config files follow the existing lint exclusions).

Draft / remaining validation: No Portkey credentials were available to record real provider responses. The repository requires real-provider e2e cassettes, pinned/latest scenario variants, paired span-tree snapshots, and CI summary entries for new instrumentation. That provider-backed scenario coverage is still outstanding; the synthetic unit-test responses above are not provider recordings.