Add explicit-state distribution fitting path

[jonathanberthias]

Current situation

The parameter graph and explicit State already exist, but the public distribution path still fits by rebuilding distributions with fitted values. Node also defines arithmetic that always creates a FunctionExpr; a distribution needs to be a graph node, but cannot validly use those numeric-expression operators.

Change

Add an opt-in continuous distribution node, state-aware likelihood evaluation, and fit_mle. A FitResult retains the model and holds fitted values in a separate state. Literal constructor values become structural constants, while supplied Parameter nodes keep their identity. Move arithmetic from Node to Expr, leaving the shared node base responsible for graph traversal.

Notes

The existing public distribution API remains in place. Distribution arithmetic can later define operations with distribution-specific results; this PR gives no meaning to Distribution + Distribution. The new path is a continuous starting point, not the eventual common contract for discrete or multivariate laws.