Grilling: NPU as the default OCR engine?

Part of #65

Blocked by: [Prototype: standalone NPU-vs-CPU OCR bench in the proto worktree](https://github.com/adityanandanx/heimdall/issues/67)

## Question

Given the prototype numbers: **does RapidOCR-on-NPU become heimdall's default OCR engine** (replacing the onnxruntime/CPU RapidOCR path in `src/heimdall/capture/ocr.py`)?

Weigh: per-frame latency vs CPU; cold-init cost (daemon start / engine warm-up); robustness over a session (watchdog/failure → auto-fallback to CPU?), recompile-on-shape-change behavior; and the CPU-usage win that motivated the question in the first place.

Resolution of this grilling = the **decision** that closes the map (Destination in #65): NPU-default, NPU-opt-in-behind-CPU, or stay CPU — stated with the tradeoffs and one-line recommendation for the follow-up implementation effort.