Replace per-chunk HTTP requests to output filters with a single
bidirectional streaming connection per filter. This eliminates
the 50-200+ round-trips per streaming LLM response.

Filters opt in via streaming: true in config. When all output filters
support streaming, brightstaff opens one POST per filter with a streaming
request body (Body::wrap_stream) and reads the streaming response. Filters
that don't opt in fall back to the existing per-chunk behavior.

Updates the PII deanonymizer demo as the reference implementation with
request.stream() + StreamingResponse support.

Made-with: Cursor