Add passive Observable nodes for interpretability

[LucaSforza]

Summary

• Keeps computational and observation graphs separate: Observable edges are compiled additively for interpretability and never alter model topology or outputs.
• Adds stereotype-driven Observable editor nodes, fixed semantic handles, validation, isolated type checking, and NNTree/compiler support.
• Adds ActivationRecorder and ActivationStatistics runtime analyses with execution-mode gating, lifecycle finalization, W&B publication, and local fallback storage.
• Preserves remote backend/job artifact boundaries and serialization guarantees: observables remain outside model layers, parameters, state, and exported wheels.
• Updates architecture, user, training, stereotype, and Observable runtime documentation.

Validation evidence

• Frontend: 325 passed, 5 skipped; pnpm --dir front-end check and production build.
• Integration smoke: 63 passed, 4 skipped.
• MCP: 45 passed.
• Backend/Python: 190 passed.
• Sphinx documentation build passed.
• Browser-derived MNIST remote end-to-end job completed successfully.