fix(streaming): resolve architectural flaws in SSE streaming

[chandan-1427]

Overview
Following a review of the recent message/stream implementation, this PR addresses 4 critical production-grade architectural issues that would prevent the streaming feature from safely scaling under load.

Architectural Fixes
Resolved Database DDoS Risk (Polling Hammer): > * Issue: The while True loop was polling the storage layer every 100ms infinitely, which would instantly max out DB connection pools under multi-user load.

Fix: Implemented an exponential backoff for poll_interval capping at 2.0s to drastically reduce storage queries while maintaining responsiveness.

Fixed Telemetry Memory Leak: > * Issue: track_active_task was incrementing bindu_active_tasks on stream creation but failing to decrement on successful completion, leading to permanently inflated active task metrics.

Fix: Added decrement logic for create_operations upon successful function exit.

Closed Protocol Security Bypass: > * Issue: The isinstance(jsonrpc_response, Response) check was too broad, intercepting any Starlette response and bypassing strict a2a_response_ta validation.

Fix: Narrowed the type check strictly to StreamingResponse.

Fixed Infinite Loading on Stream Crash: > * Issue: If the stream generator crashed (e.g. DB outage) and the subsequent "failed" state update also failed, it yielded the last known state (working).

Fix: Forced the error_event to explicitly yield state: "failed" and final: True to guarantee the client stream closes gracefully.

Testing
Ran pytest suite locally (665 passing). Verified core A2A and message handlers remain stable.