feat: sequence mode (cross-frame warm-start) + ROI/seed CLI

[johnaoga]

feat(engine): sequence mode (cross-frame warm-start) + ROI/seed CLI controls

Enables cuNCorr to track real high-strain DIC sequences, where the previous fixed-reference, per-frame path collapsed once cumulative displacement exceeded the coarse seed-search radius.

Sequence mode (the main change):

• DicEngine::process_sequence / NcorrSession::process_frame_sequence / reset_sequence, plus CLI --sequence. Each grid point initialises from the previous frame's converged warp (cross-frame seed propagation); only cold points pay a coarse search. Reuses the existing GPU icgn_batch_cuda and CPU IC-GN cores unchanged. process_frame stays stateless for parity-oracle use.
• On real coating data this holds coverage across the full 141-frame sequence (t7 frame 140: 345 -> 34,254 valid points) and matches the CppNCorr CPU reference to ~0.01 px median, while running faster than the fixed path (warm-start skips the per-frame coarse search): ~1.6 min on an A100 vs ~36 min for the 32-thread CPU pipeline.

Supporting CLI/engine controls:

• --roi (apply the ROI mask; previously the CLI ignored it and even fed the mask in as a frame), --seed X,Y (manual CPU seed, x=col/y=row), and --seed-search N (configurable coarse-search radius; was hard-wired to 15).
• SessionConfig gains seed_x/seed_y/seed_search (cuNCorr extensions).

Build fix:

• dic_engine.cpp: add missing (relied on transitive include under GCC 8.5; fails on GCC >= 13).