Add benchmarking infrastructure

[chmjkb]

Build automated benchmarking infrastructure to measure and track model performance across devices and configurations.

Currently, benchmarking models consumes a lot of time:

• Manual testing on each device
• No standardized benchmark suite
• Results scattered across different runs/people
• No comparison between model configurations (quantization, delegates)

Goals

• Automated benchmark suite
• Standardized metrics across all models
• Historical tracking to catch performance regressions