A non-asymptotic theory of Kernel Ridge Regression: deterministic equivalents, test error, and GCV estimator