ONNX Session Enhancements & Voice Features

A new single chokepoint was introduced in ONNX session management for provider selection and session construction.

CoreML acceleration is now conditionally stripped when initializer dimensions exceed the Metal 2D-texture ceiling of 16384, preventing out-of-memory errors on lower-RAM Macs.

The ONNX session builder now retries CPU-only construction if any session creation fails with a hardware accelerator.

Voice recording and playback now feature screen-awake functionality using a wake lock, ensuring long audio sessions are not prematurely terminated by the OS.

Tests were expanded to 8 real-file scenarios for ONNX sessions, covering oversized initializers and fail-open mechanisms.

The embedding service now uses a serialized queue and a single daemon worker to prevent race conditions and Out-of-Memory crashes during concurrent inference calls.

The minimum required Python version was bumped from 3.9 to 3.11 to align with actual dependency stack requirements.

Documentation and FAQ were updated to clarify the architectural boundary involving the new session helper and the CoreML texture limit.

• Single chokepoint for provider selection and session construction in ONNX backend.

• CoreML EP is stripped if any initializer dimension exceeds 16384.

• Screen wake lock implemented for voice recording and playback.

• Embedding service concurrency issues resolved using a single serialization queue.

• Minimum Python requirement raised from 3.9 to 3.11.

• Added real-file tests for ONNX session behavior across various initialization scenarios.