CodeRabbit: "Fix" Catch-up-math

[AlexanderLavelle]

From #484 --

-        for _ in range(
-            self.start_window,
-            (loaded_checkpoint_window + 1) * self.hparams.inner_steps,
-        ):
-            self.inner_scheduler.step()
+        # Advance scheduler by the exact number of inner steps since the start window
+        steps_to_apply = max(
+            0,
+            (loaded_checkpoint_window - self.start_window + 1)
+            * self.hparams.inner_steps,
+        )
+        for _ in range(steps_to_apply):
+            self.inner_scheduler.step()

self.sync_window = loaded_checkpoint_window

Code rabbit explains that one over-steps the LR scheduler by start_window * (inner_steps-1)

This is likely negligible but we should still modify the math to achieve the desired state