[BUG] Native Ollama still uses 128k fallback for context management before fetching model metadata in v3.74.0

[pmltn]

/api/show reports context_length = 1,000,000 and the webview
correctly displays 1M after PR #878.

However, after restarting VS Code or rebuilding the API handler,
an existing task condenses around 125k.

NativeOllamaHandler does not implement ensureModelFetched().
Therefore Task.safeEnsureModelFetched() is a no-op and getModel()
returns openAiModelInfoSaneDefaults with contextWindow = 128,000.

PR #1053 fixed the same lifecycle bug for RouterProvider, but
NativeOllamaHandler was not included.