perf(orchestrator): batch TelemetryStore lock acquisitions

[AlexMikhalev]

Context

Two TelemetryStore methods acquire the RwLock more often than necessary:

1. all_model_performances() acquires N+2 read locks for M models (inconsistent snapshot possible)
2. record_telemetry() acquires write lock once per event instead of batching

Proposed fix

1. Single read lock pass in all_model_performances()
2. New record_batch() method with single write lock

Refs #537 (Gitea), Refs #523

[AlexMikhalev]

Implemented and verified. Closing as part of sprint housekeeping.