Make a distinction between load errors in the test file being loaded vs. in other namespaces.

If an exception occurs while Kaocha loads a test namespace, then a warning is printed, the :kaocha.type/namespace test is marked as failed, and loading/running continues.

This is so that when you have a typo which prevents a test namespace from loading, this doesn't stop the world. Instead it just runs the rest of the tests, and reports a failure pointing at the syntax error.

The problem with this approach is that loading a namespace could fail because a dependent namespace has issues. In the degenerate case a source namespace that is used indirectly in every test fails to load, causing a failure to be reported for every test namespace. This is especially easy to do (and annoying to encounter) in --watch mode.

Instead when loading a test namespace we should first evaluate its ns form. If that causes issues then we hard abort the test run. Only after the ns form has loaded do we load the full namespace, and turn load errors into test failures.

Start by tracing Kaocha's test-namespace loading path and the :kaocha.type/namespace test behavior. Check how the namespace's ns form is evaluated separately from the remaining forms, then verify that dependency load errors abort while test-namespace load errors become failures. Done means the distinction works in normal and --watch runs.