Set model to eval mode for knn and linear classification

[guarin]

Our knn and linear classifiers used for benchmarking do not set the model to evaluation mode before training. Instead we only set requires gradients to False, see:

lightly/lightly/utils/benchmarking/knn_classifier.py

Lines 144 to 150 in 54a1b7b

	def on_fit_start(self) -> None:
	# Freeze model weights.
	deactivate_requires_grad(model=self.model)
	def on_fit_end(self) -> None:
	# Unfreeze model weights.
	activate_requires_grad(model=self.model)

The correct way to do this would be to set the model to eval mode before every training epoch:

def on_train_epoch_start(self) -> None:
    if self.freeze_model:
        self.model.eval()

Note that it has to happen in on_train_epoch_start because pytorch lightning automatically sets models to train mode at the beginning of the training epoch.

This was already implemented in #1263 but the PR is not yet merged to the main branch (and probably won't be merged for a while).