fix: handle response.incomplete when streaming

[TheGreatAxios]

Closes CL-9424

Summary

• A streamed response.incomplete event (for example max_output_tokens or a content filter) used to fall through to the default case, so the turn finished as if it had succeeded.
• It now throws a ProtocolMismatchError with the incomplete_details reason, as parseJSONResponse does for status incomplete. runInference turns it into a protocol_mismatch inference.error.
• Usage from that event is not reported. @intx/inference 0.4's harness drops the events of a parseResponse call that throws, ignores inference.error events from adapters, and has no end-of-stream hook, so an adapter can't emit usage and fail on the same event.
• New e2e test streams response.incomplete through runInference and asserts a protocol_mismatch error containing the reason and no inference.done.

Checks

bun run check, bun run build, bun run test:e2e and the Node pack smoke pass locally.