Add policyengine.derivations for per-variable computation explanations

[MaxGhenis]

Summary

Adds a new policyengine.derivations module so any caller can ask "how did PolicyEngine arrive at this value?" for a single household and get either a structured object or a plain-prose walkthrough.

from policyengine.derivations import derive, narrate

derivation = derive(simulation, "income_tax_before_credits", 2026)
print(derivation.value)                       # 3220.0
print(derivation.trace_text(max_depth=4))     # indented dependency tree
print(derivation.top_level_contributions())   # [(name, value), ...]
narrative = narrate(derivation, country="us", household_summary="TX, joint")

What's in it

• Derivation and TraceNode — frozen dataclasses for the pruned computation tree (deterministic).
• derive(simulation, variable, period) — runs the calculation with simulation.trace = True, captures the tree, returns a Derivation.
• top_level_contributions(derivation) — the immediate dependencies of the root, useful for "breakdown by component" tables.
• narrate(derivation, ...) / narrate_async(...) — optional LLM narration via LiteLLM. Lazy-imports litellm so the deterministic path has no LLM dependency.

Why this is upstream

policybench needed per-cell "PolicyEngine derivation" panels in its prediction-detail modal. The deterministic primitives are general — calculators, dashboards, and papers can use them too — so they belong in policyengine.py rather than being baked into a downstream consumer. policybench will be refactored to consume this module.

Test plan

• Hermetic tests for is_zero_value, zero-subtree pruning, and narrate/narrate_async (with a fake LiteLLM client). The PolicyEngine-US integration tests in tests/test_derivations.py exercise derive against a real one-adult household.
• ruff check + ruff format clean.

🤖 Generated with Claude Code

[MaxGhenis]

CI failure (Test 3.13/3.14) is a pre-existing upstream issue: policyengine_core's new check_computation_modes rejects policyengine_us.self_employment_income because it uses both adds and uprating. The same failure reproduces on a stale main checkout because the test workflow installs the latest published policyengine (which transitively pulls the broken combo), and it is independent of this branch's changes — the new code only adds the derivations module.

The hermetic unit tests for derivations pass; the integration test that exercises derive on a real US Simulation also passes locally with a working policyengine_us install.