Optimize runtime complexity of linter when passing in multiple snippets

[netomi]

This fixes #557 .

Instead of calculation all root at the start we delay that when evaluating each individual snippet.
The roots data structure is populated with all nodes of all imported jsonnet snippets, so we utilize that also in the importFunc where we try to resolve imports from the list of roots to avoid doing duplicate imports.

However this is a best effort only, as the mechanism is not smart, but is just able to resolve imports in the same path. This could be further improved, but for the test cases that is already sufficient, as all of them are in the same directory.

With this fix, the multiple tests run in around 0.23s compared to 9s before.