feat(cli): add provider-doctor command for tiered provider/model E2E diagnostics

Adds `jcode provider-doctor <provider>` to take any OpenAI-compatible
provider/model through the strict end-to-end checkpoints the coverage
ledger tracks, as a user-facing debugging tool rather than an
opencode-only test path.

Three tiers, escalate only as needed:
- offline: no key, no spend; validates jcode wiring (catalog reload,
  picker, fallback labeling, model switch) against a synthetic catalog
- catalog (default): needs key, ~no spend; adds live /models fetch
- full: needs key, spends balance; adds chat, streaming, tool-call loop

Only the full tier can earn strict (READY) coverage; lighter tiers
record API-dependent checkpoints as skipped so nothing is over-credited.

Human-readable output with PASS/FAIL/skip per checkpoint plus a verdict
and next-step hint on first failure; --json for scripting/CI; non-zero
exit when the chosen tier did not fully pass.

Extracts the four pure-HTTP live probe fns out of test-gated
lifecycle_driver.rs into live_provider_probes.rs so they compile into
the shipping binary. Docs in docs/PROVIDER_DOCTOR.md.