New features:
- NetworkGradient operator: expr.grad(model) for Jacobian/gradient computation
- stop_gradient support via jno.np.stop_gradient(expr)
- Stochastic Noise operator (gaussian/uniform/laplace) for SDEs
- LoRA + per-layer LR masking shown in README 2D Poisson example
- Soft BCs, integral tracker, and gradient-alignment tracker in README example
- Updated 07_analysis tutorial with in-training cosine-similarity tracker,
  sparse output-layer mask, and post-training NTK analysis
- Adds Differential-Operators documentation page

Bug fixes:
- Fix Model.params stale-reference bug after buffer donation (convert to @property)

Testing:
- Fix 45 slow smoke tests (feax/diffrax import guards, OOM reductions,
  GPU pre-allocation fix via XLA_PYTHON_CLIENT_PREALLOCATE=false)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>