perf: Optimize telemetry latency logging and feature flag caching

Optimizations implemented:

1. Latency Logger (latency_logger.py):
   - Eliminated extractor pattern - replaced wrapper classes with direct
     attribute access functions, removing object creation overhead
   - Switched from time.perf_counter() to time.monotonic() for faster timing
   - Added feature flag early exit - checks cached telemetry_enabled flag
     to skip heavy work when telemetry is disabled
   - Simplified code structure with early returns for better readability

2. Feature Flag Caching (feature_flag.py):
   - Changed cache key from session_id to host
   - Feature flags now shared across multiple connections to same host
   - Reduces network calls - first connection fetches, subsequent reuse cache

Performance impact:
- When telemetry disabled: significantly reduced overhead (only timing + debug log)
- When telemetry enabled: reduced overhead from data extraction optimizations
- Feature flag: Only fetched once per host instead of per session
- Overall: Reduces telemetry overhead substantially

The decorator now:
- Always logs latency at DEBUG level for debugging
- Exits early using cached connection.telemetry_enabled flag
- Only performs data extraction and object creation when telemetry is enabled

Author: samikshya-db

src/databricks/sql/common/feature_flag.py

@@ -165,8 +165,9 @@ def get_instance(cls, connection: "Connection") -> FeatureFlagsContext:
             cls._initialize()
             assert cls._executor is not None
 
-            # Use the unique session ID as the key
-            key = connection.get_session_id_hex()
+            # Cache at HOST level - share feature flags across connections to same host
+            # Feature flags are per-host, not per-session
+            key = connection.session.host
             if key not in cls._context_map:
                 cls._context_map[key] = FeatureFlagsContext(
                     connection, cls._executor, connection.session.http_client
@@ -177,7 +178,8 @@ def get_instance(cls, connection: "Connection") -> FeatureFlagsContext:
     def remove_instance(cls, connection: "Connection"):
         """Removes the context for a given connection and shuts down the executor if no clients remain."""
         with cls._lock:
-            key = connection.get_session_id_hex()
+            # Use host as key to match get_instance
+            key = connection.session.host
             if key in cls._context_map:
                 cls._context_map.pop(key, None)