aGrUM/pyAgrum 3.2.0 released

pyAgrum

• New: KTBN (k-order dynamic Bayesian networks) (thanks to Seth Aguila and Anis Khacef): learn, generate
  and run inference on dynamic Bayesian networks with more than one time slice of memory. Inference (filtering
  and smoothing) works over any time window with interventions AND observations, and the memory order K itself
  can be learned from data instead of being fixed by hand.
  Replaces pyagrum.lib.dynamicBN (now deprecated), with new notebooks.
• Lighter, faster import: pyAgrum's Python modules (markov_random_field, influence_diagram,
  credal_net, causal_model, prm, ktbn) are now loaded only when needed instead of all at once.
• Cleaner error messages (internal technical details no longer shown).
• Minor notebook fixes.
• CausalModel from a DAG: CausalModel can now be built from a plain DAG, without a BayesNet
  (hasObservationalBN()); structural queries work either way, CPT-dependent ones raise
  OperationNotAllowed. connectedComponents() return type aligned with other graphical models.

aGrUM

• New: KTBN (k-order dynamic Bayesian networks) (thanks to Seth Aguila and Anis Khacef): C++ engine behind
  the pyAgrum feature above -- construction, inference over an arbitrary time window (filtering and smoothing),
  automatic learning of the memory order K, and random generation.
• Windows build reliability improvements (several linking/build fixes).
• New CI job testing pyAgrum wheels against a conda-forge-like environment on Windows.
• CausalModel decoupled from BayesNet: new DAG-only constructor; observationalBN() now optional
  (hasObservationalBN() guard); fixes inducedCausalSubModel silently building a BN with
  uninitialized CPTs. connectedComponents() now returns NodeProperty<NodeId>, like DAGmodel/DiGraph.
• BUILD_SHARED_LIBS=ON support: per-module shared libraries for pyAgrum on conda-forge, with
  GUM_PUBLIC_<MODULE> visibility tagging across all modules and numerous accompanying Windows
  link/build fixes.
• Documentation improvements and typo fixes.