Revisit macOS ONNX accelerator selection for Whispering

[braden-w]

PR #1946 forced OrtAccelerator::CpuOnly on macOS for Whispering's shipped ONNX engines because our current int8 Parakeet/Moonshine setup measured faster on CPU than through OrtAccelerator::Auto/CoreML.

Measured result from that investigation:

• CPU-only: ~3.5 us/sample
• Warm CoreML through Auto: ~12 us/sample
• Cold CoreML: ~24 us/sample

Upstream data point: cjpais/transcribe-rs#86
Downstream fix: #1946

This is not currently a Whispering bug. CPU-only is the measured correct default for our shipped macOS ONNX models.

Revisit this when one of these changes:

• transcribe-rs adds model-aware provider selection
• Handy/transcribe-rs moves these models to a transcribe.cpp/ggml backend
• Whispering ships different ONNX models or quantization formats
• CoreML EP support changes enough that int8 Parakeet/Moonshine should be re-benchmarked

Done when:

• Current macOS transcription models are re-benchmarked
• The CPU-only override is removed or narrowed only if Auto/CoreML is faster for the shipped models