[codex] add neural watermark scaffold

[CEQ151]

Summary

• add a neural watermark v1 scaffold alongside the legacy blind watermark engine
• add ML codec, models, attacks, dataset, training, ONNX export, benchmarking, and traceability utilities
• wire engine selection, diagnostics, and fallback handling through the Python helper, Electron main process, preload bridge, and renderer UI
• add configs, docs, model manifest placeholders, packaging updates, and focused tests for codec and traceability

Why

This lays the project groundwork for a trainable robust image watermark path that can survive common network degradation while preserving the existing legacy extraction path and improving GitHub-level traceability for future model iterations.

Validation

• cmd /c npm run typecheck
• python -m unittest blind_watermark.tests.test_mlwm_codec blind_watermark.tests.test_mlwm_traceability
• python -m compileall blind_watermark
• python blind_watermark\bwm_helper.py --mode check --models-dir resources\models\neural_wm
• smoke-tested engine=auto fallback to legacy extraction when neural runtime/models are unavailable

Notes

• neural runtime is scaffolded but not yet active because trained ONNX models and the Python 3.12 ML environment have not been produced in this repository yet
• two pre-existing untracked PNG files in �lind_watermark/ were intentionally left untouched