(WiP) feat: Add cost-based usage limits: UsageLimits.cost_limit_usd and BudgetGuard

[OchnikBartek]

Summary

Adds cost-based usage limits in two complementary layers:

• UsageLimits.cost_limit_usd — per-run cost cap, checked after each model response. Accepts int/str/Decimal (rejects float/bool). Accumulated on RunUsage.total_cost_usd via genai-prices; fail-open when a model is unpriced.
• BudgetGuard capability — cumulative budget across runs over a rolling window, with a pluggable BudgetStore (InMemoryBudgetStore, SQLiteBudgetStore) and per-tenant key_fn.

Adds ModelResponse.cost_or_none(), docs (docs/budget.md, cost_limit_usd section in docs/agent.md, API reference), and tests.

What to review

• Is the BudgetGuard API ergonomic for the client-API-key use case? (key_fn, per-tenant buckets)
• Sensible defaults? (window_hours=24, retention_hours=168, fail_open=False)
• Does cost_limit_usd as a per-run cap cover the need?
• The provider_name == 'test' / function: short-circuit in cost_or_none — OK or smell?

Test plan

• make lint — clean
• make typecheck (pyright + mypy) — clean
• make test — 70 new tests in tests/test_budget.py + tests/test_usage_limits.py, full suite green
• make docs — builds clean