Feat/add streaming chunks based guardrails

[vrushabh-akto]

Summary

• Async prefetching pipeline for streaming guardrails — while batch N's Akto API call is in-flight, batch N+1 accumulates in parallel. Result collection is near-instant when the API responds within the batch fill window (pipeline ok), eliminating per-batch stall latency for the agent.
• Separate thread pools — hook-level checks (request/response) and stream batch checks run on independent pools so streaming load never starves prompt-level guardrails.
• Fixed stream response payload format — responsePayload is now JSON-encoded ({"content": [{"type": "text", "text": "..."}]}) matching the structure of requestPayload, fixing guardrail evaluation on stream batches.
• 403 block responses — blocked requests/responses now return 403 with {"error": "<reason>"} instead of a graceful SSE message. Mid-stream blocks send data: {"error": "..."}\n\ndata: [DONE].
• Configurable threshold — AKTO_TEXT_THRESHOLD env var (default 600 chars) controls batch size; tunable after measuring API latency from logs.
• Structured logging — latency on every Akto API call, full payload and response logging behind AKTO_LOG_PAYLOADS flag, clear REQUEST / STREAM / RESPONSE prefixes, pipeline status (pipeline ok vs waited Xms) per batch.