The previous algorithm for `getNotebookType()` was based on getting the
kernel spec out of the current notebook. This is a problem when format
is triggered by, for example, `jupyterlab-autosave-on-focus-change`,
immediately after notebook creation. In that case, the language is
determined to be `null` and errors arise.

Now the `onSave()` handler waits for the `sessionContext` to be ready,
then `getNotebookType()` gets the same kernel spec info from the session
rather than the loaded notebook.

The `!` operator is a type assertion that the attribute is not null, but
does nothing at runtime to ensure this is the case. I haven't yet found
reproduction steps for this failure, but I encountered a case where
accessing `model!.sharedModel` threw an error. This commit adds optional
chaining operators, which do provide protection at runtime.

`getNotebookType()` has two ways to determine the language. If the
metadata check fails, it falls through to the current session check.
Previously, if metadata was `null`, the method would return early.