Detect mid-region power failures via NVM in-region flag

[byeongjee]

Summary

All three algorithms (MILP, RockClimb, SCHEMATIC) assume power never fails mid-region. This PR detects violations of that assumption (e.g. from wrong energy estimates), always on — including non-debug builds.

Firmware

• New FRAM words __nvm_in_region / __nvm_violation in each runtime.
• save_pc_sp_and_halt clears __nvm_in_region once the boundary checkpoint is committed; boot code (and the wait-mode continue path) sets it when execution (re)enters a region. Cost: two 1-word FRAM writes per region.
• New check_region_violation boot macro: a reset arriving with the flag set records __nvm_violation = 1 and parks forever blinking LED1 (P1.0). The park is stable across further resets, so evidence survives until the next reflash.
• bench_halt() / debug_exit() clear the flag at completion so post-run mspdebug attach resets cannot false-positive.

Python

• New RegionViolationError; check_region_violation() / raise_if_region_violation() in device/flash.py.
• ckpt bench reads the flag after every device run — including when the Saleae capture times out, which is exactly what a parked device causes — and writes the CSV row with status region_violation.
• ckpt verify reports a violation as Status.ERROR with a "region energy violation" detail on both the normal readback and the capture-timeout path.

Testing

• All 54 boot.S variants (3 algos x 3 halt modes x debug on/off x 3 CPU freqs) and runtime C files compile; all linker combinations resolve.
• Disassembly confirms commit-point ordering: flag cleared after __nvm_pc/__nvm_sp are written, set again on region re-entry.
• ruff format/ruff check clean; 139 unit-marked pytest tests pass.
• Not yet exercised on hardware.