Prune zero polynomial zonotope terms

[MoritzMaibaum]

Motivation

• Keep PolynomialZonotope term dictionaries minimal by removing monomials whose coefficients cancel to exact zero after merging duplicate canonical exponents, so later arithmetic and propagation do not carry redundant zero terms.

Description

• Add _is_zero_coeff(value, *, atol=0.0) that detects exact-zero coefficients for torch tensors, nested fallback tuples, and scalars while allowing explicit tolerance via atol only when opt-in.
• Centralize pruning in PolynomialZonotope.__init__ by summing duplicate canonical exponent coefficients and skipping terms whose merged coefficient is exactly zero, avoiding scattered pruning across arithmetic helpers.
• Preserve certified-arithmetic semantics by using exact zero detection by default (no implicit tolerance), and support torch-backed and fallback tuple/float coefficient representations uniformly.
• Add regression tests exercising duplicate-exponent summation, cancellation pruning, removal of all-zero torch tensor coefficients, preservation of exact tiny nonzero coefficients, and constructor-mediated pruning for operations that rebuild via the constructor.

Testing

• Ran pytest -q tests/test_polynomial_zonotope.py, which passed (all tests in that file passed).
• Ran the full test suite python -m pytest -q, which completed successfully with all tests passing (158 passed, 1 warning).

---

Codex Task

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e713eb5df8

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Avoid per-term CUDA synchronization during pruning

When a PZ uses CUDA coefficients, .item() here forces a device-to-host synchronization for every term on every constructor call; since arithmetic paths rebuild PolynomialZonotope objects, large GPU-backed propagations will serialize around these checks even when no term is zero. Consider deferring/skipping pruning for non-CPU tensors or batching it so GPU work is not synchronized per monomial.

Useful? React with 👍 / 👎.