test: make models-list tests hermetic and parallelizable

[dgageot]

What

Make the TestModelsListCommand_* tests hermetic and parallelizable.

Why

Each of the five TestModelsListCommand_* tests ran newModelsCmd().Execute()
against the real environment, which made them slow (~2–3s each, ~12–16s for
the whole cmd/root package):

1. config.AvailableProviders probes ~19 providers × several env vars. For every
   missing API key, the default env-provider chain shells out to the macOS
   security keychain (plus pass/op/Docker-Desktop socket). Measured at
   ~1.48s for 20 lookups.
2. The real models.dev store fetches https://models.dev/api.json (30s timeout)
   on a cold cache — a network dependency in a unit test.
3. They used t.Setenv + paths.SetConfigDir, so they could not run in parallel.

How

• Add a small modelsCmdOption variadic seam to newModelsCmd / newModelsListCmd.
  Production calls newModelsCmd() with no args, so behavior is unchanged.
  Tests inject a hermetic env provider (environment.NewMapEnvProvider) and an
  in-memory models.dev store (modelsdev.NewDatabaseStore) via the existing
  RuntimeConfig.EnvProviderForTests / RuntimeConfig.ModelsDevStoreOverride
  fields.
• Rewrite models_test.go: drop t.Setenv / paths.SetConfigDir, add
  t.Parallel() to every test, and add a tiny in-memory testCatalog() database.
• The catalog includes a catalog-only model so the tests actually verify the
  injected store is read (not just the per-provider defaults); the default-marker
  test now asserts exactly one default row matching the auto-selected model; and
  the canceled-context test asserts the server is never reached.

Result

• Each models test: ~2–3s → 0.00s
• cmd/root package: ~12–16s → ~1.7s (~3.7s with -race)
• Validation: golangci-lint 0 issues, go test -race ./cmd/root/ green,
  full suite passes with no regressions.