fix: intermittent app insights event loss in serverless env

[cristian-groza]

Summary

• Register atexit flush handler on every track_event call (not just CLI), so buffered events are sent on graceful shutdown even if the caller never explicitly flushes
• Parse and use IngestionEndpoint from the connection string instead of always hitting the global default endpoint
• Replace SynchronousSender with _DiagnosticSender that logs HTTP 400 (silently discarded by SDK), network timeouts, and other failures at WARNING level
• Log remaining queue size after flush to surface partial send failures

Context

AgentRun.Start custom events were intermittently missing from App Insights despite serverless logs confirming the SDK track_event call completed. Root causes:

1. Events buffered but never flushed — the agent execution path didn't register an atexit handler (only CLI did), so if the process exited before the finally block, the in-memory queue was lost
2. SDK silently discards HTTP 400 — SenderBase.send() swallows 400 responses without logging, making schema/ikey rejections invisible
3. Wrong ingestion endpoint — _parse_connection_string() only extracted InstrumentationKey, ignoring IngestionEndpoint, causing events to route to the global endpoint instead of the regional one