Onboarding: skip escape hatch + Ollama connect coverage

[TheGreatAxios]

Summary

• Investigated the report that onboarding's "Bring your own AI" step has no local/self-hosted option and no way past it. On origin/main, apps/web/src/onboarding.ts / onboarding-page.tsx already carry full Ollama support (CL-6146: URL-shaped card, http://localhost:11434 default, /v1 base-URL normalization) — the live repro this ticket describes likely hit a deploy that predates that landing. This PR closes the two gaps that were still real:
• Adds end-to-end frontend coverage for the Ollama card: selecting it prefills the default URL, submitting sends it as baseURL with the fixed Ollama placeholder key, and the tenant lands on the workbench with a usable model. Previously only the backend (packages/onboarding/test/complete-credential.test.ts) covered this.
• Adds a "Skip for now" control to the credential step — an account with no provider ready yet (URL unreachable, no key in hand) can now proceed to / instead of being blocked. Provisioning already treats this as bench_unseeded, an anticipated state; the no-usable-model banner from CL-6568: tell people before they type, and don't lie when it fails #280 (CL-6568) is what tells the person, honestly, that a connection still needs finishing.

Test plan

• bun test test/onboarding.test.tsx — 49 pass, including the two new tests
• bunx tsc --noEmit on apps/web — clean
• bunx eslint on changed files — no new warnings (one pre-existing, unrelated warning on onboarding-page.tsx)
• bunx prettier --check on changed files — clean
• bun test (full apps/web suite) — 846 pass / 3 fail, all 3 pre-existing and unrelated (network-dependent tests hitting ECONNREFUSED, and a tailwind-build test that needs a prior bun run build)

Linear: CL-6576