nnR 0.2.0

New features

Added Pnm() for the general neural-network polynomials in Definition 2.27 of Rafi, Padgett, and Nakarmi (2024).
Added column-batch evaluation to inst() and reusable realizations through realize_nn().
Added stk_many() and nn_sum_many() for efficient n-ary algebraic operations. Unequal-depth realization identities are documented as ReLU-specific because their padding uses tunnel networks.
Generalized MC() from one-dimensional inputs to the full multidimensional construction in Definition 2.39, including the valid one-sample case.
Added validate_nn() and repaired draw_nn().

is_nn() now verifies numeric, finite, dimensionally connected layers.
Sum(n, k) now supports every positive n and k, as its definition requires.
Prd(q, eps) now allocates the square-network error budget using eps / (2^(q - 1) + 1).
Approximation scaling is computed in the log domain and extreme parameters now fail with a direct floating-point-range error instead of producing non-finite network weights.
Composition reports incompatible end widths directly.
Infix algebraic operators are now explicitly exported, while implementation helpers remain internal.
The previously auto-exported implementation objects A, B, A_k, C_k, and ck are no longer public. Use Phi() or Phi_k() instead.

Power networks are constructed iteratively without duplicate recursion.
Taylor polynomials reuse a single power sequence and n-ary sum.
Activations and realization are vectorized over complete matrices.
Copy, sum, identity, tunnel, norm, maximum, and maximum-convolution constructors avoid repeated matrix growth.
Removed package-internal source() calls and duplicate infix implementations.
Added an R CMD check workflow and modernized artifact upload for the paper workflow.