feat: reproducible cost benchmark + provider base-URL override

[prashar32]

Two things, shipped together because the second is what the first needs.

1. Reproducible cost benchmark (benchmark/)

The flagship "dollars saved" measurement. The same looping agent runs twice against a deterministic mock provider — once ungoverned, once through RiskKernel with a hard dollar budget — and the spend is compared, read straight from RiskKernel's own ledger:

baseline (no governance)  50 calls  $0.6250
governed (RiskKernel)     20 calls  $0.2500   ← halted: dollar_budget_exceeded
dollars saved             $0.3750  (60%)

Key-free and reproducible (python3 benchmark/benchmark.py): the mock returns fixed token usage and pricing.json pins the model price, so the only variable is whether the budget stopped the loop. The governed spend is measured, not modelled (GET /v1/runs/{id} → usage.dollars). The cap is flat no matter how long the runaway would have continued — the saving grows without bound; see the README methodology.

2. Provider base-URL override (the enabler)

RISKKERNEL_OPENAI_BASE_URL / RISKKERNEL_ANTHROPIC_BASE_URL point a native provider at a gateway, proxy, or local mock instead of the default endpoint — a chainable WithBaseURL(...), wired from config. RiskKernel-namespaced so it never collides with the caller-facing OPENAI_BASE_URL used to point an app at RiskKernel (which would make RiskKernel forward to itself in a shared shell).

Verified

WithBaseURL trims/defaults correctly and a Chat call routes to the override (asserted against a mock); config reads the namespaced vars; go test ./... green (13 packages). The benchmark runs end-to-end and halts the governed loop at exactly the budget.