feat: add structured training telemetry schema and validation pipeline hooks

[VarshiniGunti]

Summary

This PR introduces a structured telemetry foundation for logs-to-training workflows in oan-ai-api, aligned with agentic training requirements. It standardizes how chat and tool-use events are captured so downstream pipelines can reliably build training datasets for supervised fine-tuning and preference optimization.

What This PR Adds

• A versioned training event schema (v1) for:
  • user
  • assistant
  • tool_call
  • tool_result
  • error
• Runtime event emission from the chat flow in a machine-readable format:
  • training_event=<json>
• Event validation utilities to enforce:
  • valid tool names (against registered tools)
  • required tool_call_id on tool events
  • correct tool_call → tool_result linkage
• Documentation for schema and usage
• A CLI validator to verify extracted JSONL/log lines before downstream processing

Implementation Details

• Added telemetry models/builders and emitter utilities in app/telemetry/
• Instrumented app/services/chat.py to emit:
  • user event at request start
  • tool call/result events from model message parts
  • assistant event after stream completion
  • error event on exceptions
• Added helpers/validate_training_events.py for offline validation
• Documented schema and usage in:
  • docs/training_pipeline/log_event_schema.md
  • README.md

Why This Change Matters

This creates a consistent, auditable event contract for both Q&A and multi-step agentic interactions. It improves trace quality for training data generation and introduces guardrails to catch malformed or inconsistent tool trajectories early.

Validation

Executed:

python -m py_compile app/telemetry/events.py app/telemetry/validator.py app/services/chat.py helpers/validate_training_events.py

Compilation and syntax checks passed successfully.